Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
gaoqiong
composable_kernel_ROCM
Commits
f23a2e2a
Commit
f23a2e2a
authored
Feb 11, 2025
by
Jakub Piasecki
Browse files
resolved conflicts
parents
f3eb5a18
c0adab48
Changes
340
Hide whitespace changes
Inline
Side-by-side
Showing
20 changed files
with
289 additions
and
110 deletions
+289
-110
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
...c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
+18
-6
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
...c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
+18
-6
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
...c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
+18
-6
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
...c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
+17
-5
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
...lu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
+17
-5
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
...lu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
+17
-5
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
...lu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
+17
-5
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
...lu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
+17
-5
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
...d_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
+11
-3
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
...d_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
+11
-4
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
...d_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
+11
-4
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
...d_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
+11
-3
library/src/tensor_operation_instance/gpu/gemm_b_scale/device_gemm_b_scale_xdl_f16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn.hpp
...16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn.hpp
+34
-38
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
..._fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
+15
-3
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
..._fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
+15
-3
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
..._fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
+15
-3
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
..._fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
+15
-3
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp
...xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp
+4
-1
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp
...xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp
+4
-1
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp
...xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp
+4
-1
No files found.
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -38,23 +38,30 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[k, m], b[k, n], d0[m, n], d1[m, n]
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
128
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
128
,
128
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
2
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -111,18 +118,22 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
2
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -134,7 +145,8 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -38,23 +38,30 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[k, m], b[n, k], d0[m, n], d1[m, n]
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
128
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
128
,
128
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
8
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -111,18 +118,22 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
8
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -134,7 +145,8 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -38,23 +38,30 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
128
,
128
,
128
,
64
,
16
,
2
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
128
,
128
,
128
,
32
,
8
,
2
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
8
,
2
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -111,18 +118,22 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
2
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -134,7 +145,8 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -38,23 +38,30 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[m, k], b[n, k], d0[m, n], d1[m ,n]
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
128
,
128
,
128
,
64
,
16
,
16
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif
// clang-format on
>
;
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -102,18 +109,22 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -125,7 +136,8 @@ using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddAddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -23,22 +23,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[k, m], b[k, n], d0[m, n]
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
2
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -95,18 +102,22 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instanc
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
2
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -118,7 +129,8 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregul
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -23,22 +23,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[k, m], b[n, k], d0[m, n], d1[m, n]
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
8
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -95,18 +102,22 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instanc
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
8
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -118,7 +129,8 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregul
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -23,22 +23,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
8
,
2
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -95,18 +102,22 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instanc
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
2
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -118,7 +129,8 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregul
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -23,22 +23,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[m, k], b[n, k], d0[m, n], d1[m ,n]
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -86,18 +93,22 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instanc
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -109,7 +120,8 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregul
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_Tuple
,
F16
,
PassThrough
,
PassThrough
,
AddFastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -28,8 +28,10 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
AddReluAdd
=
ck
::
tensor_operation
::
element_wise
::
AddReluAdd
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmMNKPadding
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKPadding
;
#endif
// e = elementwise((a * b), d0, d1)
// h = layernorm(e, gamma, beta)
...
...
@@ -37,11 +39,13 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
template
<
LoopScheduler
GemmLoopScheduler
,
PipelineVersion
GemmPipeline
>
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
2
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
...
...
@@ -58,17 +62,20 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instan
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
128
,
64
,
32
,
8
,
8
,
32
,
32
,
2
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
2
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
// pipeline v1, 1 wave
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
...
...
@@ -81,7 +88,8 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_irregu
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -28,20 +28,23 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
AddReluAdd
=
ck
::
tensor_operation
::
element_wise
::
AddReluAdd
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmMNKPadding
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKPadding
;
#endif
// e = elementwise((a * b), d0, d1)
// h = layernorm(e, gamma, beta)
// outout: h[m, n]
// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
template
<
LoopScheduler
GemmLoopScheduler
,
PipelineVersion
GemmPipeline
>
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
8
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
...
...
@@ -58,17 +61,20 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instan
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
128
,
64
,
32
,
8
,
8
,
32
,
32
,
2
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
8
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
// pipeline v1, 1 wave
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
1
,
1
,
S
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
...
...
@@ -81,7 +87,8 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_irregu
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Col
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
1
,
1
,
S
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -28,20 +28,23 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
AddReluAdd
=
ck
::
tensor_operation
::
element_wise
::
AddReluAdd
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmMNKPadding
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKPadding
;
#endif
// e = elementwise((a * b), d0, d1)
// h = layernorm(e, gamma, beta)
// outout: h[m, n]
// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
template
<
LoopScheduler
GemmLoopScheduler
,
PipelineVersion
GemmPipeline
>
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
8
,
2
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
...
...
@@ -58,17 +61,20 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instan
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
2
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
// pipeline v1, 1 wave
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
...
...
@@ -81,7 +87,8 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_irregu
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Row
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -28,8 +28,10 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
AddReluAdd
=
ck
::
tensor_operation
::
element_wise
::
AddReluAdd
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmMNKPadding
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKPadding
;
#endif
// e = elementwise((a * b), d0, d1)
// h = layernorm(e, gamma, beta)
...
...
@@ -37,11 +39,13 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
template
<
LoopScheduler
GemmLoopScheduler
,
PipelineVersion
GemmPipeline
>
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
32
,
8
>
,
8
,
S
<
32
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
16
,
8
>
,
8
,
S
<
16
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
...
...
@@ -55,17 +59,20 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instan
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
16
,
8
>
,
8
,
S
<
16
,
8
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
16
,
4
>
,
8
,
S
<
16
,
4
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmDefault
,
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
<
16
,
4
>
,
8
,
S
<
16
,
4
>
,
1
,
GemmLoopScheduler
,
GemmPipeline
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| A| B| Ds| H| AData| BData| AccData| CShuffle| DsData| EMeanVarData| GammaData| BetaData| HData| A| B| CDE| H| 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| PostShuffle| PostShuffle| Layernorm| Layernorm| LoopScheduler| Pipeline|
//#######################################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Type| Type| Type| Elementwise| 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| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize| | |
//#######################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _M_N| _NWaveNPerXdl| _M_N| _M| | |
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
// pipeline v1, 1 wave
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
...
...
@@ -78,7 +85,8 @@ using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_irregu
,
DeviceGemmMultipleDLayernorm_Xdl_CShuffle
<
Row
,
Col
,
Row_Row_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
F16_F16_Tuple
,
F16
,
F16
,
F16
,
F16
,
PassThrough
,
PassThrough
,
AddReluAdd
,
PassThrough
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
<
16
,
4
>
,
1
,
S
<
16
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances
(
...
...
library/src/tensor_operation_instance/gpu/gemm_b_scale/device_gemm_b_scale_xdl_f16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn.hpp
View file @
f23a2e2a
...
...
@@ -48,55 +48,51 @@ using device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_instances = std::tuple<
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//Compute friendly
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
128
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
16
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
128
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
16
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
128
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
16
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//0
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//1
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//3
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//4
//Latency friendly
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
4
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
32
,
128
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//5
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
4
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//6
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//7
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
32
,
128
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//8
// Memory friendly v3
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
128
,
32
,
128
,
8
,
32
,
32
,
32
,
2
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
128
,
16
,
128
,
8
,
16
,
16
,
16
,
4
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
32
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
16
,
128
,
8
,
16
,
16
,
16
,
2
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
4
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
4
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
32
,
128
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
64
,
128
,
8
,
32
,
16
,
16
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
64
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
128
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
128
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
16
,
256
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
32
,
256
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
128
,
32
,
128
,
8
,
32
,
32
,
32
,
2
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//9
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
128
,
16
,
128
,
8
,
16
,
16
,
16
,
4
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//10
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
32
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//11
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
16
,
128
,
8
,
16
,
16
,
16
,
2
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//12
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//13
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
32
,
128
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//16
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
64
,
128
,
8
,
32
,
16
,
16
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//17
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
64
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//18
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
128
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//19
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
128
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//20
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
16
,
256
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//21
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
32
,
256
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//22
// Memory friendly v4
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
32
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
16
,
128
,
8
,
16
,
16
,
16
,
2
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
4
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
4
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
32
,
128
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
64
,
128
,
8
,
32
,
16
,
16
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
64
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
128
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
128
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
16
,
256
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
32
,
256
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
32
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//23
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
64
,
16
,
128
,
8
,
16
,
16
,
16
,
2
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//24
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
2
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//25
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
1
,
128
,
16
,
16
,
128
,
8
,
16
,
16
,
16
,
1
,
1
,
S
<
16
,
4
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//26
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
32
,
128
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//28
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
64
,
128
,
8
,
32
,
16
,
16
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//29
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
64
,
128
,
8
,
32
,
32
,
32
,
1
,
1
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//30
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
16
,
128
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//31
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
1
,
128
,
32
,
128
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//32
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
16
,
256
,
128
,
8
,
32
,
16
,
16
,
1
,
4
,
S
<
16
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
4
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//33
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
32
,
256
,
128
,
8
,
32
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v4
,
half_t
,
half_t
,
false
,
false
>
,
//34
//new Compute friendly kernel
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
4
,
1
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
2
,
2
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//35
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
128
,
128
,
64
,
8
,
32
,
32
,
32
,
4
,
1
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
,
//36
//new Memory friendly kernel
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
16
,
64
,
256
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
32
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
DeviceGemm_Xdl_CShuffleV3
<
Row
,
Col
,
Row
,
F16
,
I4
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
1
,
128
,
16
,
64
,
256
,
8
,
32
,
16
,
16
,
1
,
1
,
S
<
32
,
8
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
32
,
32
,
0
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
BlkGemmPipeSched
,
BlockGemmPipelineVersion
::
v3
,
half_t
,
half_t
,
false
,
false
>
//37
// clang-format on
>
;
}
// namespace instance
...
...
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -22,22 +22,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// outout: e[m, n]
// input: a[k, m], b[k, n]
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
2
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -94,17 +101,21 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::t
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
2
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -116,6 +127,7 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_ins
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -22,22 +22,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// outout: e[m, n]
// input: a[k, m], b[k, n]
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
2
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
2
,
8
,
32
,
32
,
2
,
4
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -94,17 +101,21 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::t
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
2
,
8
,
32
,
32
,
1
,
2
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -116,6 +127,7 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_ins
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Col
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -22,22 +22,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// outout: e[m, n]
// input: a[m, k], b[k, n]
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif
// clang-format on
>
;
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
8
,
2
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
8
,
2
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -94,17 +101,21 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::t
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
64
,
128
,
32
,
8
,
2
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -116,6 +127,7 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_ins
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Row
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
View file @
f23a2e2a
...
...
@@ -22,22 +22,29 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// outout: e[m, n]
// input: a[m, k], b[n, k]
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_generic_instance
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif
// clang-format on
>
;
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
1
,
256
,
128
,
128
,
64
,
16
,
16
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -85,17 +92,21 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::t
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmDefault
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
// irregular tile size
using
device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_irregular_tile_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################| A| B| Ds| E| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| 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| | |
//##############################| | | | | | | | | | | 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
64
,
16
,
16
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -107,6 +118,7 @@ using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_irregular_tile_ins
,
DeviceGemmMultipleD_Xdl_CShuffle
<
Row
,
Col
,
Empty_Tuple
,
Row
,
F16
,
F16
,
F32
,
F32
,
Empty_Tuple
,
F16
,
PassThrough
,
PassThrough
,
FastGelu
,
GemmMNKPadding
,
1
,
64
,
16
,
16
,
32
,
8
,
8
,
16
,
16
,
1
,
1
,
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
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp
View file @
f23a2e2a
...
...
@@ -33,11 +33,13 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
template
<
ck
::
tensor_operation
::
device
::
GemmSpecialization
GemmSpec
>
using
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_iw_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
//PipelineVersion::v1; interwave
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
2
,
8
,
true
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
4
,
8
,
true
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
...
...
@@ -57,6 +59,7 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_iw_instances = std::tuple<
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
128
,
32
,
4
,
8
,
32
,
32
,
2
,
1
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
1
,
8
,
true
,
1
,
1
,
S
<
1
,
32
,
1
,
4
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
32
,
32
,
4
,
8
,
32
,
32
,
1
,
1
,
S
<
1
,
2
,
32
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
2
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
16
,
32
,
4
,
8
,
16
,
16
,
1
,
2
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
2
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
4
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp
View file @
f23a2e2a
...
...
@@ -34,11 +34,13 @@ template <ck::tensor_operation::device::GemmSpecialization GemmSpec,
ck
::
PipelineVersion
PipVer
,
ck
::
LoopScheduler
LoopSche
>
using
device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
16
,
128
,
4
,
8
,
16
,
16
,
1
,
4
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
4
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
F16
,
PipVer
,
LoopSche
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
16
,
256
,
4
,
8
,
16
,
16
,
1
,
8
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
8
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
4
,
F16
,
PipVer
,
LoopSche
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
16
,
256
,
4
,
8
,
16
,
16
,
1
,
4
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
4
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
16
>
,
4
,
F16
,
PipVer
,
LoopSche
>
,
...
...
@@ -63,6 +65,7 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances = std::tup
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
128
,
16
,
8
,
8
,
16
,
16
,
4
,
1
,
S
<
1
,
8
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
8
,
16
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
1
,
8
,
true
,
1
,
1
,
S
<
1
,
32
,
1
,
4
>
,
4
,
F16
,
PipVer
,
LoopSche
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
256
,
16
,
8
,
8
,
16
,
16
,
8
,
1
,
S
<
1
,
8
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
8
,
16
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
1
,
8
,
true
,
1
,
1
,
S
<
1
,
32
,
1
,
4
>
,
4
,
F16
,
PipVer
,
LoopSche
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
256
,
16
,
8
,
8
,
16
,
16
,
4
,
1
,
S
<
1
,
8
,
32
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
8
,
16
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
2
,
1
,
8
,
true
,
1
,
1
,
S
<
1
,
64
,
1
,
4
>
,
4
,
F16
,
PipVer
,
LoopSche
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp
View file @
f23a2e2a
...
...
@@ -33,12 +33,14 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
template
<
ck
::
tensor_operation
::
device
::
GemmSpecialization
GemmSpec
>
using
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_iw_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//PipelineVersion::v1; interwave
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
3
,
8
,
8
,
true
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
64
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
3
,
8
,
8
,
true
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
3
,
8
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
...
...
@@ -52,6 +54,7 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_iw_instances = std::tuple<
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
128
,
32
,
128
,
4
,
8
,
32
,
32
,
1
,
2
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
32
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
3
,
8
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
64
,
32
,
4
,
8
,
32
,
32
,
2
,
1
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
3
,
8
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
,
DeviceGemmXdlSplitKCShuffle
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
64
,
32
,
64
,
4
,
8
,
32
,
32
,
1
,
2
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
2
,
1
,
3
>
,
S
<
0
,
2
,
1
,
3
>
,
3
,
8
,
8
,
true
,
S
<
1
,
4
,
16
,
1
>
,
S
<
0
,
1
,
3
,
2
>
,
S
<
0
,
1
,
3
,
2
>
,
3
,
8
,
8
,
true
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
8
,
F16
,
PipelineVersion
::
v1
,
LoopScheduler
::
Interwave
>
#endif // !defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
Prev
1
…
11
12
13
14
15
16
17
Next
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment
