ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-13 09:45:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add support for double buffering in direct load GEMM kernel (#1052)

Browse Source 
This PR introduces support for double buffering in LDS into GEMM kernels that use direct load instructions.

Direct loads now use inline asm instead of intrinsics. Usage of intrinsics results in compiler adding additional waitcnt instructions what breaks possible load/compute overlap in case of double buffering.

Usage of inline asm results in the need to use sched_barrier in order to make sure that compiler cannot incorrectly reschedule instructions since it does not know the data dependencies between global->LDS and LDS->registers.
This commit is contained in:
Bartlomiej Wroblewski
2023-12-03 23:08:47 +01:00
committed by GitHub
parent c7d5c7727b
commit bc4bf9bd03
10 changed files with 211 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp
View File

@@ -35,7 +35,21 @@ using device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances =
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmMNPadding, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 32, 32, 64, 8, 8, 32, 32, 1, 1, S<1, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<1, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<1, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<1, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<2, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<2, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 0, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 0, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 32, 128, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 64, 32, 32, 64, 8, 8, 32, 32, 1, 1, S<1, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<1, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 128, 64, 32, 32, 8, 8, 32, 32, 1, 1, S<2, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<2, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmMNPadding, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F16, F16, F16, F32, F32, PassThrough, PassThrough, PassThrough, GemmMNPadding, 2, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, S<4, 16, 4>, S<1, 0, 2>, 2, 2, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;

3
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp
View File

@@ -32,7 +32,8 @@ using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instances =
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;

3
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp
View File

@@ -32,7 +32,8 @@ using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instances =
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Col, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;

3
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp
View File

@@ -31,7 +31,8 @@ using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instances =
// ##################################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Row, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<0, 2, 1>, 1, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;

5
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp
View File

@@ -24,8 +24,7 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances = std::tuple<
// clang-format off
@@ -34,7 +33,7 @@ using device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instances =
// ##################################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmMNPadding, 1, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
DeviceGemm_Xdl_CShuffle_LdsDirectLoad< Row, Col, Row, F32, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, GemmDefault, 2, 256, 64, 64, 32, 8, 8, 32, 32, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, 1, 1, S<1, 8, 1, 8>, 4>
// clang-format on
>;