Skip to content

GitLab

"...git@developer.sourcefind.cn:cnjsdfcy/simbricks.git" did not exist on "913f172f71ec3e6e47e8916add7a81082023bce3"
Commit 3ac4aea4 authored by j4yan's avatar j4yan
Browse files

adding more tuning parameters 

add tuning parameters for other 3 layouts
parent 2ee18035
Show whitespace changes
Inline Side-by-side
Showing with 106 additions and 7 deletions
include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r3.hpp
View file @ 3ac4aea4
...@@ -175,6 +175,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B ...@@ -175,6 +175,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
"wrong!"); "wrong!");
// TODO: remove this restriction // TODO: remove this restriction
static_assert(BM0 == 2, "wrong");
static_assert(BM0 == 2 && BN0 == 2, "wrong"); static_assert(BM0 == 2 && BN0 == 2, "wrong");
} }
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dlops_f32_f32_f32_km_kn_mn_instance.cpp
View file @ 3ac4aea4
...@@ -24,13 +24,47 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa ...@@ -24,13 +24,47 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// Compilation parameters for a[k, m] * b[k, n] = c[m, n] // Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_dlops_f32_f32_f32_km_kn_mn_instances = std::tuple< using device_gemm_dlops_f32_f32_f32_km_kn_mn_instances = std::tuple<
// clang-format off // clang-format off
// ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| // ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | // ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> /*
// DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 4, 1, S<8, 1>, S<8, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> * K1 = 1
*/
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 1, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 1, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 1, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 1, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 1, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
/*
* K1 = 2
*/
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 2, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 2, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 2, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 2, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 2, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// repeat the above configurartion, but changing K1 to 4, NOT working for fp32
// DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 4>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<4, 1, 1, 4>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 4, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 4>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
// DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 4, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 4>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
// DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 4, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 4>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
// DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 4, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 4>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
// DeviceGemmDlops< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 4, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 4>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 4>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 4>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on // clang-format on
>; >;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dlops_f32_f32_f32_km_nk_mn_instance.cpp
View file @ 3ac4aea4
...@@ -29,8 +29,29 @@ using device_gemm_dlops_f32_f32_f32_km_nk_mn_instances = std::tuple< ...@@ -29,8 +29,29 @@ using device_gemm_dlops_f32_f32_f32_km_nk_mn_instances = std::tuple<
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | // ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> /*
// DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 4, 1, S<8, 1>, S<8, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> * K1 = 1
*/
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 1, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 1, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 1, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 1, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 1, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
/*
* K1 = 2
*/
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 2, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 2, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 2, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 2, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 2, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on // clang-format on
>; >;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dlops_f32_f32_f32_mk_kn_mn_instance.cpp
View file @ 3ac4aea4
...@@ -29,8 +29,29 @@ using device_gemm_dlops_f32_f32_f32_mk_kn_mn_instances = std::tuple< ...@@ -29,8 +29,29 @@ using device_gemm_dlops_f32_f32_f32_mk_kn_mn_instances = std::tuple<
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | // ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> /*
// DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 4, 1, S<8, 1>, S<8, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> * K1 = 1
*/
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 1, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 1, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 1, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 1, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 1, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
/*
* K1 = 2
*/
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 2, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 2, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 2, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 2, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 2, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on // clang-format on
>; >;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dlops_f32_f32_f32_mk_nk_mn_instance.cpp
View file @ 3ac4aea4
...@@ -29,7 +29,29 @@ using device_gemm_dlops_f32_f32_f32_mk_nk_mn_instances = std::tuple< ...@@ -29,7 +29,29 @@ using device_gemm_dlops_f32_f32_f32_mk_nk_mn_instances = std::tuple<
// ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ##########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | | // ##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // ##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> /*
* K1 = 1
*/
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 1, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 1, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 1, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 1, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 1, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 1, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<4, 1, 4, 1>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
/*
* K1 = 2
*/
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 8, 2, 8, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 8, 2, 4, 8, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 8, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 8, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 8, 1, S<8, 2>, S<4, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 128, 8, 2, 8, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 128, 64, 8, 2, 8, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>,
DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 128, 8, 2, 4, 8, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 4, 1, S<8, 1>, S<8, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> // DeviceGemmDlops< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 8, 2, 4, 4, 1, S<8, 1>, S<8, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on // clang-format on
>; >;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment