Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "9cae3c4fb9e342b7c2b039f0853f1e74e58d4bbe"
Unverified Commit 3786bfe1 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

added padding of K into gemm_v2r3 (#887) 



* added kpad support into v2r3

* add generic instances

* fixed comments

* fixed mnk padding

* Update device_batched_gemm_xdl.hpp

---------
Co-authored-by: default avatarJing Zhang <jizha@amd.com>
parent a61b8b78
Show whitespace changes
Inline Side-by-side
Showing with 72 additions and 11 deletions
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
View file @ 3786bfe1
......@@ -185,7 +185,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpecialization::MNPadding,
GemmSpecialization::MNKPadding,
MPerBlock,
NPerBlock,
K0PerBlock,
......@@ -315,11 +315,6 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
return false;
}
if(problem.K % K1 != 0)
{
return false;
}
return GridwiseGemm::CheckValidity(problem);
}
......@@ -416,7 +411,12 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerXDL << ", "
<< NPerXDL << ", "
<< MXdlPerWave << ", "
<< NXdlPerWave << ", "
<< ">"
<< " NumGemmKPrefetchStage: "
<< NumGemmKPrefetchStage << ", "
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ 3786bfe1
......@@ -194,7 +194,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
StrideC{StrideC_},
MPadded{CalculateMPadded(M_)},
NPadded{CalculateNPadded(N_)},
K0{CalculateK0(K)}
K0{CalculateK0(K_)}
{
}
......@@ -383,7 +383,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
__host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / (K0PerBlock * K1);
const index_t num_loop = math::integer_divide_ceil(K, K0PerBlock * K1);
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
......@@ -840,7 +840,25 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
}
}();
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{
const auto K0Pad = math::integer_divide_ceil(K0, K0PerBlock) * K0PerBlock;
const auto KPad = K0Pad * K1Value;
const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(K0Pad, K1Value)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
return transform_tensor_descriptor(
a_grid_desc_m_k,
......@@ -874,7 +892,26 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
}
}();
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{
const auto K0Pad = math::integer_divide_ceil(K0, K0PerBlock) * K0PerBlock;
const auto KPad = K0Pad * K1Value;
const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(K0Pad, K1Value)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
return transform_tensor_descriptor(
b_grid_desc_k_n,
......
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
View file @ 3786bfe1
......@@ -25,6 +25,16 @@ 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
//#################| 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 | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
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>
// 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
......@@ -100,6 +110,8 @@ void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(
DeviceBatchedGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_generic_instances{});
add_device_operation_instances(instances,
device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances{});
}
......
library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
View file @ 3786bfe1
......@@ -25,6 +25,16 @@ 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
//#################| 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 | | |
//#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
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>
// 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
......@@ -88,6 +98,8 @@ void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances, device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_generic_instances{});
add_device_operation_instances(instances,
device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances{});
}
......
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