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
9ba504b6
Commit
9ba504b6
authored
Feb 07, 2025
by
ThomasNing
Browse files
merge with the develop support the fp8 with computev4
parents
e3402c93
f49de496
Changes
198
Hide whitespace changes
Inline
Side-by-side
Showing
20 changed files
with
177 additions
and
41 deletions
+177
-41
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
...ice_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
+5
-1
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
...ice_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
+10
-2
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
...ice_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
+10
-2
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
...xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+12
-4
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
...cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
+7
-2
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
...xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+7
-2
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
...cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
+7
-2
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
...xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+7
-2
library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
..._conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
+20
-4
library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
...d_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
+20
-4
library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
...l_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
+20
-4
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
...m_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
+5
-1
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
...e_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
+5
-0
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
...vice_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
+5
-1
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
...vice_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
+5
-1
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
...vice_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
+8
-2
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
No files found.
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
View file @
9ba504b6
...
...
@@ -27,12 +27,15 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using
device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
2
,
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
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -65,6 +68,7 @@ using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v2
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Col
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
View file @
9ba504b6
...
...
@@ -26,23 +26,30 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_generic_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
64
,
16
,
16
,
4
,
16
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
64
,
16
,
16
,
4
,
8
,
16
,
16
,
1
,
1
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif
// clang-format on
>
;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using
device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -102,6 +109,7 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v2
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Row
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
View file @
9ba504b6
...
...
@@ -26,23 +26,30 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_generic_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
64
,
32
,
64
,
4
,
16
,
32
,
32
,
1
,
2
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
64
,
32
,
64
,
4
,
8
,
32
,
32
,
1
,
2
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
true
,
S
<
4
,
16
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
8
,
true
,
7
,
1
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#endif
// clang-format on
>
;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using
device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK| LoopScheduler| Pipeline|
//#################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch| | |
//#################| | | | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| Stage | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -90,6 +97,7 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v2
>
,
DeviceBatchedGemmXdl
<
F16
,
F16
,
F16
,
F32
,
Row
,
Col
,
Row
,
PassThrough
,
PassThrough
,
PassThrough
,
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
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @
9ba504b6
...
...
@@ -26,18 +26,22 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
Scale
=
ck
::
tensor_operation
::
element_wise
::
Scale
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmPadded
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKOPadding
;
#endif
// c[g, m, n] = a[g, m, k] * b[g, n, k]
template
<
bool
Masking
>
using
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskOut|
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
64
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
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
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
256
,
256
,
128
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
4
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
1
,
256
,
128
,
256
,
32
,
64
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
8
,
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
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
...
...
@@ -53,24 +57,28 @@ using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
128
,
128
,
64
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
128
,
64
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
template
<
bool
Masking
>
using
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskOut|
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
256
,
128
,
40
,
64
,
32
,
4
,
4
,
2
,
32
,
32
,
2
,
4
,
2
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
256
,
128
,
40
,
128
,
32
,
4
,
4
,
2
,
32
,
32
,
2
,
4
,
4
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
128
,
256
,
40
,
64
,
32
,
4
,
4
,
2
,
32
,
32
,
1
,
8
,
2
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
128
,
256
,
40
,
128
,
32
,
4
,
4
,
2
,
32
,
32
,
1
,
8
,
4
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
128
,
128
,
40
,
64
,
32
,
4
,
4
,
2
,
32
,
32
,
1
,
4
,
2
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
1
,
256
,
128
,
128
,
40
,
128
,
32
,
4
,
4
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
2
,
128
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
4
,
4
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
Masking
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance
(
...
...
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
View file @
9ba504b6
...
...
@@ -26,10 +26,12 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
ScaleAdd
=
ck
::
tensor_operation
::
element_wise
::
ScaleAdd
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmPadded
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKOPadding
;
static
constexpr
auto
TensorDefault
=
ck
::
tensor_operation
::
device
::
TensorSpecialization
::
Default
;
#endif
// c[g, m, n] = a[g, m, k] * b[g, n, k]
template
<
index_t
NumDimG
,
...
...
@@ -40,11 +42,13 @@ template <index_t NumDimG,
MaskingSpecialization
MaskingSpec
>
using
device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
// #############################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskingSpec| D0s Bias|
// #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | SrcScalar|
// #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | PerVector|
// #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<
BF16
>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
64
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
,
1
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<
BF16
>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
64
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
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
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<
BF16
>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
4
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
...
...
@@ -62,7 +66,8 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<
BF16
>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
128
,
64
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
,
1
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<
BF16
>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
128
,
64
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<
BF16
>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
64
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances
(
...
...
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @
9ba504b6
...
...
@@ -26,10 +26,12 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
ScaleAdd
=
ck
::
tensor_operation
::
element_wise
::
ScaleAdd
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmPadded
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKOPadding
;
static
constexpr
auto
TensorDefault
=
ck
::
tensor_operation
::
device
::
TensorSpecialization
::
Default
;
#endif
// c[g, m, n] = a[g, m, k] * b[g, n, k]
template
<
index_t
NumDimG
,
...
...
@@ -40,11 +42,13 @@ template <index_t NumDimG,
MaskingSpecialization
MaskingSpec
>
using
device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
// #############################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskingSpec| D0s Bias|
// #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | SrcScalar|
// #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | PerVector|
// #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<
F16
>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
64
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
,
1
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<
F16
>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
64
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
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
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<
F16
>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
4
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
...
...
@@ -64,7 +68,8 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<
F16
>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
128
,
64
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
,
1
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<
F16
>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
128
,
64
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<
F16
>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
ScaleAdd
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
64
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances
(
...
...
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
View file @
9ba504b6
...
...
@@ -26,10 +26,12 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
Scale
=
ck
::
tensor_operation
::
element_wise
::
Scale
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmPadded
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKOPadding
;
static
constexpr
auto
TensorDefault
=
ck
::
tensor_operation
::
device
::
TensorSpecialization
::
Default
;
#endif
// c[g, m, n] = a[g, m, k] * b[g, n, k]
template
<
index_t
NumDimG
,
...
...
@@ -40,11 +42,13 @@ template <index_t NumDimG,
MaskingSpecialization
MaskingSpec
>
using
device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
// #############################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskingSpec|
// #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| |
// #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| |
// #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
64
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
64
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
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
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
4
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
...
...
@@ -60,7 +64,8 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
64
,
256
,
64
,
64
,
32
,
8
,
8
,
2
,
16
,
16
,
1
,
16
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
4
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
BF16
,
BF16
,
BF16
,
BF16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
128
,
64
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances
(
...
...
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @
9ba504b6
...
...
@@ -26,10 +26,12 @@ using S = ck::Sequence<Is...>;
using
PassThrough
=
ck
::
tensor_operation
::
element_wise
::
PassThrough
;
using
Scale
=
ck
::
tensor_operation
::
element_wise
::
Scale
;
#if !defined(CK_USE_AMD_MFMA_GFX950)
static
constexpr
auto
GemmDefault
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
Default
;
static
constexpr
auto
GemmPadded
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKOPadding
;
static
constexpr
auto
TensorDefault
=
ck
::
tensor_operation
::
device
::
TensorSpecialization
::
Default
;
#endif
// c[g, m, n] = a[g, m, k] * b[g, n, k]
template
<
index_t
NumDimG
,
...
...
@@ -40,11 +42,13 @@ template <index_t NumDimG,
MaskingSpecialization
MaskingSpec
>
using
device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
// #############################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskingSpec|
// #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| |
// #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| |
// #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
#else
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
64
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
64
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
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
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
256
,
128
,
32
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
2
,
4
,
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
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
...
...
@@ -62,7 +66,8 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
64
,
256
,
64
,
64
,
32
,
8
,
8
,
2
,
16
,
16
,
1
,
16
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
true
,
S
<
16
,
16
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
4
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
<
NumDimG
,
NumDimM
,
NumDimN
,
NumDimK
,
NumDimO
,
F16
,
F16
,
F16
,
F16
,
ck
::
Tuple
<>
,
ck
::
Tuple
<>
,
F32
,
F16
,
PassThrough
,
PassThrough
,
Scale
,
PassThrough
,
PassThrough
,
GemmPadded
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
TensorDefault
,
1
,
256
,
128
,
128
,
64
,
128
,
32
,
8
,
8
,
2
,
32
,
32
,
1
,
4
,
4
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
false
,
S
<
8
,
32
,
1
>
,
S
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
4
,
2
,
false
,
1
,
2
,
S
<
1
,
32
,
1
,
8
>
,
8
,
MaskingSpec
>
// clang-format on
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
void
add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances
(
...
...
library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
View file @
9ba504b6
...
...
@@ -40,11 +40,12 @@ static constexpr auto ConvFwdOddC =
// arbitrary conv
using
device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -58,16 +59,20 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances = std::tuple<
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdDefault
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// 1x1, pad 0
using
device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -81,16 +86,20 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std:
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1P0
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// 1x1, stride 1, pad 0
using
device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -104,15 +113,19 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = s
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwd1x1S1P0
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
using
device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdOddC
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdOddC
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdOddC
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
4
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
4
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -131,6 +144,9 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances = std::
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdOddC
,
256
,
256
,
64
,
2
,
4
,
32
,
32
,
4
,
1
,
S
<
2
,
32
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
32
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdOddC
,
128
,
128
,
64
,
2
,
4
,
32
,
32
,
2
,
2
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdOddC
,
128
,
64
,
64
,
2
,
4
,
32
,
32
,
1
,
2
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
PassThrough
,
ConvFwdOddC
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
View file @
9ba504b6
...
...
@@ -39,11 +39,12 @@ static constexpr auto ConvFwdOddC =
// arbitrary conv
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -57,16 +58,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances = s
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdDefault
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// 1x1, pad 0
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -80,16 +85,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instan
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1P0
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// 1x1, stride 1, pad 0
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -103,16 +112,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_ins
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwd1x1S1P0
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// Odd C
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdOddC
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdOddC
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdOddC
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
4
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
4
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -131,6 +144,9 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instanc
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdOddC
,
256
,
256
,
64
,
2
,
4
,
32
,
32
,
4
,
1
,
S
<
2
,
32
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
32
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdOddC
,
128
,
128
,
64
,
2
,
4
,
32
,
32
,
2
,
2
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdOddC
,
128
,
64
,
64
,
2
,
4
,
32
,
32
,
1
,
2
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddRelu
,
MemorySet
,
ConvFwdOddC
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
View file @
9ba504b6
...
...
@@ -37,11 +37,12 @@ static constexpr auto ConvFwdOddC =
// arbitrary conv
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -55,16 +56,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdDefault
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdDefault
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// 1x1, pad 0
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -78,16 +83,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_in
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1P0
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// 1x1, stride 1, pad 0
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -101,16 +110,20 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1S1P0
,
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
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwd1x1S1P0
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
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
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
// Odd C
using
device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_instances
=
std
::
tuple
<
// clang-format off
// clang-format off
//##############################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//##############################################################################################| | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//##############################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if !defined(CK_USE_AMD_MFMA_GFX950)
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdOddC
,
256
,
256
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdOddC
,
256
,
128
,
256
,
4
,
8
,
32
,
32
,
2
,
4
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdOddC
,
128
,
128
,
128
,
4
,
8
,
32
,
32
,
4
,
2
,
S
<
4
,
4
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
4
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
8
>
,
8
>
,
...
...
@@ -129,6 +142,9 @@ using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_ins
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdOddC
,
256
,
256
,
64
,
2
,
4
,
32
,
32
,
4
,
1
,
S
<
2
,
32
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
32
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdOddC
,
128
,
128
,
64
,
2
,
4
,
32
,
32
,
2
,
2
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
4
>
,
8
>
,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdOddC
,
128
,
64
,
64
,
2
,
4
,
32
,
32
,
1
,
2
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
2
,
16
,
4
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
16
,
1
,
1
,
4
>
,
8
>
#else
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<
F16
,
F16
,
F16
,
F32
,
PassThrough
,
PassThrough
,
AddReluAdd
,
ConvFwdOddC
,
256
,
128
,
128
,
4
,
16
,
32
,
32
,
2
,
2
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
S
<
4
,
8
,
8
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
1
,
1
,
true
,
1
,
1
,
S
<
1
,
1
,
32
,
1
,
1
,
8
>
,
8
>
#endif
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
View file @
9ba504b6
...
...
@@ -29,12 +29,15 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using
device_gemm_xdl_c_shuffle_2_stage_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| | Version|
//#####################| | | | | | | | | 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)
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
2
,
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
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
2
,
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
,
GemmDefault
,
2
,
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
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
2
,
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
>
,
...
...
@@ -82,6 +85,7 @@ using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tu
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmDefault
,
2
,
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
,
GemmDefault
,
2
,
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
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
View file @
9ba504b6
...
...
@@ -55,6 +55,11 @@ using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
BF16
,
BF16
,
BF16
,
F32
,
BF16
,
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
::
v1
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
BF16
,
BF16
,
BF16
,
F32
,
BF16
,
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
::
v1
>
,
DeviceGemm_Xdl_CShuffleV2
<
Row
,
Col
,
Row
,
BF16
,
BF16
,
BF16
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
2
,
256
,
256
,
256
,
32
,
8
,
8
,
32
,
32
,
4
,
4
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
S
<
4
,
64
,
1
>
,
S
<
1
,
0
,
2
>
,
S
<
1
,
0
,
2
>
,
2
,
8
,
8
,
0
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#if defined(CK_USE_AMD_MFMA_GFX950)
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Col
,
Row
,
BF16
,
BF16
,
BF16
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
,
BF16
,
BF16
,
BF16
,
F32
,
BF16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
#endif
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
// pipeline v1, 2 waves
,
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
View file @
9ba504b6
...
...
@@ -31,12 +31,15 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using
device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_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| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemm_Xdl_CShuffle
<
Col
,
Row
,
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
<
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
DeviceGemm_Xdl_CShuffle
<
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemm_Xdl_CShuffle
<
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemm_Xdl_CShuffle
<
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -93,6 +96,7 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle
<
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemm_Xdl_CShuffle
<
Col
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // !defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
View file @
9ba504b6
...
...
@@ -31,12 +31,15 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using
device_gemm_xdl_c_shuffle_f16_f16_f16_km_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| | |
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// pipeline v1, 1 wave
#if defined(CK_USE_AMD_MFMA_GFX950)
DeviceGemm_Xdl_CShuffle
<
Col
,
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
<
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
DeviceGemm_Xdl_CShuffle
<
Col
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemm_Xdl_CShuffle
<
Col
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
DeviceGemm_Xdl_CShuffle
<
Col
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -93,6 +96,7 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle
<
Col
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
,
DeviceGemm_Xdl_CShuffle
<
Col
,
Col
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
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
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v2
>
#endif
#endif // !defined(CK_USE_AMD_MFMA_GFX950)
// clang-format on
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
View file @
9ba504b6
...
...
@@ -35,22 +35,27 @@ static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecializati
static
constexpr
auto
MNKPadding
=
ck
::
tensor_operation
::
device
::
GemmSpecialization
::
MNKPadding
;
using
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_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
,
Row
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
1
,
8
,
1
,
1
,
1
,
S
<
1
,
16
,
1
,
8
>
,
1
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
#else
DeviceGemm_Xdl_CShuffle
<
Row
,
Row
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
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
>
;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
template
<
ck
::
tensor_operation
::
device
::
GemmSpecialization
GemmSpec
>
using
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_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
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Row
,
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
<
0
,
2
,
1
>
,
S
<
0
,
2
,
1
>
,
1
,
2
,
8
,
1
,
1
,
1
,
S
<
1
,
32
,
1
,
8
>
,
8
,
LoopScheduler
::
Default
,
PipelineVersion
::
v1
>
,
...
...
@@ -109,6 +114,7 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
DeviceGemm_Xdl_CShuffle
<
Row
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
,
DeviceGemm_Xdl_CShuffle
<
Row
,
Row
,
Row
,
F16
,
F16
,
F16
,
F32
,
F16
,
PassThrough
,
PassThrough
,
PassThrough
,
GemmSpec
,
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
>
;
...
...
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
View file @
9ba504b6
...
...
@@ -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 @
9ba504b6
...
...
@@ -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 @
9ba504b6
...
...
@@ -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 @
9ba504b6
...
...
@@ -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
>
;
...
...
Prev
1
…
3
4
5
6
7
8
9
10
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
