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
2cc0fa26
Unverified
Commit
2cc0fa26
authored
Dec 10, 2024
by
jefyang1
Committed by
GitHub
Dec 10, 2024
Browse files
Merge pull request #248 from ROCm/gfx950_mfma
Gfx950 mfma
parents
cde0f480
5008a960
Changes
64
Hide whitespace changes
Inline
Side-by-side
Showing
20 changed files
with
209 additions
and
65 deletions
+209
-65
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
...vice_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
+10
-3
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
.../device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
+5
-0
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
...l_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
+6
-4
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
...f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
+3
-0
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
...l_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
+6
-4
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
...f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
+3
-0
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
...l_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
+6
-4
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
...f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
+3
-0
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
...l_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
+3
-0
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
...f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
+3
-0
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
No files found.
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
View file @
2cc0fa26
...
...
@@ -34,23 +34,30 @@ static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_generic_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| 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|
//#####################| | | | Type| Type| Type| Type| DataType| 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| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
//DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, 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>,
//DeviceGemm_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 128, 32, 32, 16, 16, 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>, 4, LoopScheduler::Default, PipelineVersion::v1>
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
MNKPadding
,
1
,
128
,
128
,
128
,
64
,
16
,
16
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
MNKPadding
,
1
,
128
,
128
,
128
,
32
,
8
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
template
<
ck
::
tensor_operation
::
device
::
GemmSpecialization
GemmSpec
>
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| 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|
//#####################| | | | Type| Type| Type| Type| DataType| 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| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
// pipeline v1, 1 wave
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
,
...
...
@@ -100,6 +107,7 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
;
...
...
@@ -110,7 +118,6 @@ void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
{
add_device_operation_instances
(
instances
,
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_generic_instances
{});
add_device_operation_instances
(
instances
,
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances
<
GemmDefault
>
{});
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
View file @
2cc0fa26
...
...
@@ -47,6 +47,11 @@ using device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances =
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
int8_t
,
int8_t
,
int8_t
,
int32_t
,
int32_t
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
128
,
32
,
128
,
64
,
16
,
16
,
32
,
32
,
1
,
2
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
1
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
4
>
,
16
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
int8_t
,
int8_t
,
int8_t
,
int32_t
,
int32_t
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
64
,
64
,
32
,
64
,
16
,
16
,
32
,
32
,
2
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
2
>
,
16
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
int8_t
,
int8_t
,
int8_t
,
int32_t
,
int32_t
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
64
,
32
,
64
,
64
,
16
,
16
,
32
,
32
,
1
,
2
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
2
>
,
16
>
#if defined(CK_USE_AMD_MFMA_GFX950)
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
int8_t
,
int8_t
,
int8_t
,
int32_t
,
int32_t
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
256
,
128
,
128
,
128
,
32
,
32
,
32
,
32
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
1
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
1
,
1
,
1
,
S
<
1
,
64
,
1
,
4
>
,
16
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
int8_t
,
int8_t
,
int8_t
,
int32_t
,
int32_t
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
256
,
64
,
64
,
256
,
64
,
64
,
16
,
16
,
2
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
1
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
16
,
16
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
4
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -9,8 +9,7 @@ namespace device {
namespace
instance
{
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using
Instances
=
std
::
tuple
<
using
Instances
=
std
::
tuple
<
// clang-format off
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -18,6 +17,8 @@ using Instances =
//##########| 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| SrcDstVectorDim| DstScalar| | | |
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
...
...
@@ -26,9 +27,10 @@ using Instances =
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
64
,
128
,
4
,
8
,
32
,
32
,
2
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
128
,
64
,
4
,
8
,
32
,
32
,
2
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
64
,
128
,
4
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
// clang-format on
>
;
void
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_interwave_pipeline_v1_instances
(
OwnerList
<
InstanceNT
>&
instances
)
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -17,7 +17,10 @@ using Instances = std::tuple<
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmMNPadding
,
64
,
16
,
16
,
4
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -9,8 +9,7 @@ namespace device {
namespace
instance
{
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using
Instances
=
std
::
tuple
<
using
Instances
=
std
::
tuple
<
// clang-format off
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -18,6 +17,8 @@ using Instances =
//##########| 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| SrcDstVectorDim| DstScalar| | | |
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
...
...
@@ -26,9 +27,10 @@ using Instances =
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
64
,
128
,
4
,
8
,
32
,
32
,
2
,
2
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
128
,
64
,
4
,
8
,
32
,
32
,
2
,
1
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
64
,
128
,
4
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
// clang-format on
>
;
void
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_interwave_pipeline_v1_instances
(
OwnerList
<
InstanceNN
>&
instances
)
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -17,7 +17,10 @@ using Instances = std::tuple<
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmMNPadding
,
64
,
16
,
16
,
4
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -9,8 +9,7 @@ namespace device {
namespace
instance
{
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using
Instances
=
std
::
tuple
<
using
Instances
=
std
::
tuple
<
// clang-format off
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
...
...
@@ -18,6 +17,8 @@ using Instances =
//##########| 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| SrcDstVectorDim| DstScalar| | | |
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
...
...
@@ -35,9 +36,10 @@ using Instances =
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
16
,
64
,
4
,
8
,
16
,
16
,
1
,
2
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
16
,
32
,
4
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
64
,
16
,
16
,
4
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
// clang-format on
>
;
void
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_interwave_pipeline_v1_instances
(
OwnerList
<
InstanceTT
>&
instances
)
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -17,7 +17,10 @@ using Instances = std::tuple<
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmMNPadding
,
64
,
16
,
16
,
4
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -17,6 +17,8 @@ using Instances = std::tuple<
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
...
...
@@ -30,6 +32,7 @@ using Instances = std::tuple<
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
128
,
32
,
128
,
4
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
64
,
64
,
32
,
4
,
8
,
32
,
32
,
2
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
64
,
32
,
64
,
4
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
View file @
2cc0fa26
...
...
@@ -17,9 +17,12 @@ using Instances = std::tuple<
//###########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| | | |
//###########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| | | |
//###########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmMNPadding
,
64
,
64
,
32
,
4
,
8
,
32
,
32
,
2
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmMNPadding
,
256
,
128
,
144
,
8
,
8
,
16
,
16
,
2
,
9
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
8
,
8
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
2
,
2
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
,
DeviceGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmMNPadding
,
256
,
128
,
144
,
4
,
8
,
16
,
16
,
2
,
9
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
4
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
2
,
2
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Interwave
,
PipelineVersion
::
v1
>
#endif // defined(CK_USE_AMD_MFMA_GFX950)
#endif
// clang-format on
>
;
...
...
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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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 @
2cc0fa26
...
...
@@ -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
(
...
...
Prev
1
2
3
4
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
