ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-30 03:37:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

clang-format

Browse Source
This commit is contained in:
mtgu0705
2025-06-30 01:26:23 -05:00
parent 20a8b31fd6
commit 3288da013c
16 changed files with 386 additions and 359 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
include/ck/library/utility/host_tensor.hpp
View File

@@ -202,7 +202,7 @@ struct joinable_thread : std::thread
{
}
joinable_thread(joinable_thread&&) = default;
joinable_thread(joinable_thread&&) = default;
joinable_thread& operator=(joinable_thread&&) = default;
~joinable_thread()
@@ -320,7 +320,7 @@ struct Tensor
~Tensor() = default;
Tensor& operator=(const Tensor&) = default;
Tensor& operator=(Tensor&&) = default;
Tensor& operator=(Tensor&&) = default;
template <typename FromT>
explicit Tensor(const Tensor<FromT>& other) : Tensor(other.template CopyAsType<T>())

6
include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_direct_load.hpp
View File

@@ -258,7 +258,8 @@ struct ThreadGroupTensorSliceTransfer_DirectLoad
src_buf.template DirectCopyToLds<remove_cvref_t<decltype(dst_buf)>, ScalarPerVector>(
dst_buf, src_offset, dst_offset, is_src_valid);
constexpr auto move_on_dim = [&]() constexpr {
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
@@ -270,7 +271,8 @@ struct ThreadGroupTensorSliceTransfer_DirectLoad
});
return move_on_dim_;
}();
}
();
// Decide whether to move forward or backward.
constexpr auto forward_sweep = [&]() {

116
include/ck/tensor_operation/gpu/device/impl/device_moe_gemm_blockscale.hpp
View File

@@ -100,64 +100,64 @@ struct DeviceMoeGemmBlockScale
{
static constexpr index_t NumDTensor = DsDataType::Size();
using GridwiseGemm = GridwiseMoeGemmBlockScale<
ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
GemmAccDataType,
CShuffleDataType,
DsDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
BlockSize,
ScaleBlockM,
ScaleBlockN,
ScaleBlockK,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEShuffleBlockTransferScalarPerVectors,
BlkGemmPipeSched,
BlkGemmPipelineVer,
ActivationOP,
NSwizzle,
IsInputGemm,
MulRoutedWeight,
IndexType,
ComputeTypeA,
ComputeTypeB,
LDSTypeA,
LDSTypeB>;
ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
GemmAccDataType,
CShuffleDataType,
DsDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
BlockSize,
ScaleBlockM,
ScaleBlockN,
ScaleBlockK,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEShuffleBlockTransferScalarPerVectors,
BlkGemmPipeSched,
BlkGemmPipelineVer,
ActivationOP,
NSwizzle,
IsInputGemm,
MulRoutedWeight,
IndexType,
ComputeTypeA,
ComputeTypeB,
LDSTypeA,
LDSTypeB>;
using Argument = typename GridwiseGemm::Argument;

72
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp
View File

@@ -38,25 +38,25 @@ template <typename GridwiseGemm,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_multiple_d_xdl_cshuffle_lds_direct_load(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
kernel_gemm_multiple_d_xdl_cshuffle_lds_direct_load(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx90a__) || defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx90a__) || defined(__gfx94__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -73,18 +73,18 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
e_grid_desc_mblock_mperblock_nblock_nperblock,
block_2_etile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = block_2_etile_map;
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = block_2_etile_map;
#endif
}
@@ -814,16 +814,18 @@ struct GridwiseGemmMultipleD_Xdl_CShuffle_LdsDirectLoad
// A tuple of reference to C/Ds tensor descriptors.
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// A tuple of reference to C/Ds grid buffers.
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// A tuple of starting index of C/Ds blockwise copy.
const auto idx_c_ds_block_begin = container_concat(

8
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp
View File

@@ -30,12 +30,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg);
@@ -58,12 +58,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_2lds(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy
__shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];

36
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp
View File

@@ -33,12 +33,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_multi_d(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
@@ -65,12 +65,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_multi_d_2lds(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
// Pass two lds pointer is the key to tell compiler that ds_read/write
// operate on different lds chunk at same time without order dependecy
__shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -1566,16 +1566,18 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(
@@ -2031,16 +2033,18 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(

36
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp
View File

@@ -33,13 +33,13 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Even>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle(
typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
@@ -66,13 +66,13 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Even>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle_2lds(
typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
__shared__ char p_shared1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -1446,16 +1446,18 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(
@@ -1946,16 +1948,18 @@ struct GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(

4
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx_bpreshuffle.hpp
View File

@@ -34,7 +34,7 @@ template <bool Use2LDS,
TailNumber TailNum = TailNumber::Full>
__global__ enable_if_t<!Use2LDS, void>
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_mx(typename GridwiseGemm::Argument karg)
@@ -66,7 +66,7 @@ template <bool Use2LDS,
TailNumber TailNum = TailNumber::Full>
__global__ enable_if_t<Use2LDS, void>
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_gemm_xdl_cshuffle_v3_mx(typename GridwiseGemm::Argument karg)

172
include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
View File

@@ -40,12 +40,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Even>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_moe_gemm(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
@@ -75,12 +75,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Even>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_moe_gemm_2lds(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
__shared__ char p_shared1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -1714,16 +1714,18 @@ struct GridwiseMoeGemm
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin =
@@ -1744,40 +1746,40 @@ struct GridwiseMoeGemm
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
constexpr index_t scatter_weight_idx = 3; // hack fix felix
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3_scatter<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
@@ -2433,16 +2435,18 @@ struct GridwiseMoeGemm
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin =
@@ -2463,40 +2467,40 @@ struct GridwiseMoeGemm
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
constexpr index_t scatter_weight_idx = 3; // hack fix felix
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3_scatter<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());

172
include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm_blockscale.hpp
View File

@@ -40,12 +40,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Even>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_moe_gemm(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
@@ -77,12 +77,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Even>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_moe_gemm_2lds(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
__shared__ char p_shared1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -1764,16 +1764,18 @@ struct GridwiseMoeGemmBlockScale
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin =
@@ -1794,40 +1796,40 @@ struct GridwiseMoeGemmBlockScale
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
constexpr index_t scatter_weight_idx = IsInputGemm ? 1 : 1; // hack fix felix
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3_scatter<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
@@ -2504,16 +2506,18 @@ struct GridwiseMoeGemmBlockScale
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin =
@@ -2534,40 +2538,40 @@ struct GridwiseMoeGemmBlockScale
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
constexpr index_t scatter_weight_idx = IsInputGemm ? 1 : 1; // hack fix felix
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3_scatter<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());

90
include/ck/tensor_operation/gpu/grid/gridwise_moe_mx_gemm.hpp
View File

@@ -48,7 +48,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_moe_mxgemm(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
@@ -81,12 +81,12 @@ template <typename GridwiseGemm,
TailNumber TailNum = TailNumber::Even>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
// __attribute__((amdgpu_waves_per_eu(1, 1)))
kernel_moe_mxgemm_2lds(typename GridwiseGemm::Argument karg)
{
#if (!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
__shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
@@ -2797,16 +2797,18 @@ struct GridwiseMoeGemmMXBNS
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie([&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin =
@@ -2827,41 +2829,41 @@ struct GridwiseMoeGemmMXBNS
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
constexpr index_t scatter_weight_idx = 3; // hack fix felix
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3_scatter<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
// Sequence support
// arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
// Sequence support
// arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferCluster,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>, // ThreadTransferDstResetCoordinateAfterRunFlags
IndexType,
1, // ScatterDim
true, // OutputScatter: false, only use scatter weights
scatter_weight_idx // ScatterWeightIdx: ascale
>{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(0, 0, block_n_id, 0)),
c_element_op};
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());

12
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
View File

@@ -823,7 +823,8 @@ struct ThreadwiseTensorSliceTransfer_v3
buffer_(Number<buffer_offset>{}) = src_tmp_vector.template AsType<SrcData>()[i];
});
constexpr auto move_on_dim = [&]() constexpr {
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
@@ -836,7 +837,8 @@ struct ThreadwiseTensorSliceTransfer_v3
});
return move_on_dim_;
}();
}
();
// move
static_for<0, nDim, 1>{}([&](auto i) {
@@ -981,7 +983,8 @@ struct ThreadwiseTensorSliceTransfer_v3
is_dst_valid,
dst_tmp_vector.template AsType<dst_vector_t>()[Number<0>{}]);
constexpr auto move_on_dim = [&]() constexpr {
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
@@ -994,7 +997,8 @@ struct ThreadwiseTensorSliceTransfer_v3
});
return move_on_dim_;
}();
}
();
// move
static_for<0, nDim, 1>{}([&](auto i) {

2
library/include/ck/library/reference_tensor_operation/cpu/reference_moe_gemm1_blockscale.hpp
View File

@@ -110,7 +110,7 @@ struct ReferenceMoeGemm1BlockScale : public device::BaseOperator
#if CK_USE_PK4_LAYOUT_SHUFFLE
v_a = i4_to_f32_gfx9(i4);
#else
v_a = i4 - 8;
v_a = i4 - 8;
#endif
}
else

3
library/include/ck/library/reference_tensor_operation/cpu/reference_moe_gemm2_blockscale.hpp
View File

@@ -235,7 +235,8 @@ struct ReferenceMoeGemm2BlockScale : public device::BaseOperator
{0b100, +0.2500f},
{0b101, +0.3125f},
{0b110, +0.3750f},
{0b111, +0.4375f}};
{ 0b111,
+0.4375f }};
return u[i4];
}

4
library/include/ck/library/tensor_operation_instance/gpu/gemm_blockscale_wp.hpp
View File

@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#if (defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
void add_device_gemm_blockscale_wp_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_BlockScale_BPreshuffle<Row,
Col,
@@ -143,7 +143,7 @@ struct DeviceOperationInstanceFactory<
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if (defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<A0DataType, f8_t> && is_same_v<B0DataType, f8_t> &&
is_same_v<CDataType, bhalf_t>)
{

8
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_wp.hpp
View File

@@ -17,7 +17,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
#if (defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
using TGemmMulMulF8F8F16Instances =
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
@@ -240,7 +240,7 @@ void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma16
TGemmMulMulF8F8F16Instances& instances);
#endif
#if (defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
using TGemmMulMulF8F8BF16Instances =
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
Col,
@@ -347,7 +347,7 @@ struct DeviceOperationInstanceFactory<
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
// TODO: Add MFMA layout into tensor layout
#if (defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_F16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
{
@@ -422,7 +422,7 @@ struct DeviceOperationInstanceFactory<
}
#endif
#if (defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, bhalf_t>)
{