Skip to content

GitLab

Unverified Commit 2a30cfdd authored by arai713's avatar arai713 Committed by GitHub
Browse files

Merge branch 'develop' into codegen-enable-hiprtc

parents 9533a172 78195ccc
Hide whitespace changes
Inline Side-by-side
Showing with 1432 additions and 74 deletions
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -56,8 +56,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
View file @ 2a30cfdd
......@@ -74,8 +74,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -60,8 +60,7 @@ __global__ void
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -108,7 +107,7 @@ __global__ void
ignore = block_2_ctile_map;
ignore = batch_count;
ignore = compute_base_ptr_of_batch;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
#endif // end of if (defined(__gfx9__))
}
// Computes C = A * B0 * B1
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
View file @ 2a30cfdd
......@@ -83,8 +83,7 @@ __global__ void
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -68,8 +68,7 @@ __global__ void
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp
View file @ 2a30cfdd
......@@ -41,12 +41,15 @@ __global__ void
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t g_idx = blockIdx.z % karg.Batch;
const index_t k_idx = blockIdx.z / karg.Batch;
const auto a_batch_offset = karg.compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const auto b_batch_offset = karg.compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const auto ds_batch_offset = karg.compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
const auto c_batch_offset = karg.compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, k_idx);
// populate pointer, desc for Ds
static_for<0, GridwiseGemm::NumDTensor, 1>{}([&](auto i) {
// D pointer
......@@ -54,8 +57,8 @@ __global__ void
});
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + a_batch_offset,
karg.p_b_grid + b_batch_offset,
karg.p_a_grid + a_batch_offset + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + b_batch_offset + splitk_batch_offset.b_k_split_offset,
karg.p_ds_grid,
karg.p_c_grid + c_batch_offset,
p_shared,
......@@ -87,12 +90,15 @@ __global__ void
__shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t g_idx = blockIdx.z % karg.Batch;
const index_t k_idx = blockIdx.z / karg.Batch;
const auto a_batch_offset = karg.compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const auto b_batch_offset = karg.compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const auto ds_batch_offset = karg.compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
const auto c_batch_offset = karg.compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, k_idx);
// populate pointer, desc for Ds
static_for<0, GridwiseGemm::NumDTensor, 1>{}([&](auto i) {
// D pointer
......@@ -100,8 +106,8 @@ __global__ void
});
GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + a_batch_offset,
karg.p_b_grid + b_batch_offset,
karg.p_a_grid + a_batch_offset + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + b_batch_offset + splitk_batch_offset.b_k_split_offset,
karg.p_ds_grid,
karg.p_c_grid + c_batch_offset,
p_shared_0,
......@@ -303,7 +309,8 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
index_t Batch_,
AElementwiseOperation a_element_op_,
BElementwiseOperation b_element_op_,
CElementwiseOperation c_element_op_)
CElementwiseOperation c_element_op_,
index_t KBatch_)
: GridwiseGemm::Argument{p_a_grid_,
p_b_grid_,
p_ds_grid_,
......@@ -315,7 +322,7 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
StrideB_,
StrideDs_,
StrideE_,
1,
KBatch_,
a_element_op_,
b_element_op_,
c_element_op_},
......@@ -336,13 +343,14 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
arg.Print();
}
if(!GridwiseGemm::CheckValidity(arg) || arg.KBatch > 1)
if(!GridwiseGemm::CheckValidity(arg))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
index_t gdx, gdy, gdz;
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.Batch);
std::tie(gdx, gdy, gdz) =
GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.Batch * arg.KBatch);
float ave_time = 0;
......@@ -387,10 +395,11 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
rotating_mem.Next();
// clear c mem
if(arg_.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
0,
arg_.M * arg_.N * sizeof(CDataType),
stream_config.stream_id_));
hipGetErrorString(
hipMemsetAsync(arg_.p_c_grid,
0,
arg.Batch * arg_.M * arg_.N * sizeof(CDataType),
stream_config.stream_id_));
};
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
......@@ -889,7 +898,8 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
index_t BatchStrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
CElementwiseOperation c_element_op,
index_t KBatch = 1)
{
return Argument{static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......@@ -909,7 +919,8 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
Batch,
a_element_op,
b_element_op,
c_element_op};
c_element_op,
KBatch};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -934,7 +945,8 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
index_t BatchStrideE,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op) override
CElementwiseOperation c_element_op,
index_t KBatch = 1) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......@@ -954,7 +966,8 @@ struct DeviceBatchedGemmMultiD_Xdl_CShuffle_V3
Batch,
a_element_op,
b_element_op,
c_element_op);
c_element_op,
KBatch);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -59,8 +59,7 @@ __global__ void
const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -67,8 +67,7 @@ __global__ void
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -127,7 +126,7 @@ __global__ void
ignore = batch_count;
ignore = compute_base_ptr_of_batch;
ignore = c0_matrix_mask;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
#endif // end of if (defined(__gfx9__))
}
// Computes C = A * B0 * B1
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -64,8 +64,7 @@ __global__ void
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -114,7 +113,7 @@ __global__ void
ignore = batch_count;
ignore = compute_base_ptr_of_batch;
ignore = c0_matrix_mask;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
#endif // end of if (defined(__gfx9__))
}
// Computes C = A * B0 * B1
......@@ -616,6 +615,96 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
}
#ifndef __HIPCC_RTC__
static constexpr bool
IsSupported(index_t MRaw_, index_t NRaw_, index_t KRaw_, index_t Gemm1NRaw_)
{
// check vector load/store
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
// check vector load of A
if constexpr(is_same_v<ALayout, Row>)
{
if(KRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<ALayout, Col>)
{
if(MRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector load of B
if constexpr(is_same_v<BLayout, Row>)
{
if(NRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<BLayout, Col>)
{
if(KRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector load of B1
if constexpr(is_same_v<B1Layout, Row>)
{
if(Gemm1NRaw_ % B1BlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<B1Layout, Col>)
{
if(NRaw_ % B1BlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector load of C
if constexpr(is_same_v<CLayout, Row>)
{
if(Gemm1NRaw_ % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
return false;
}
}
else if constexpr(is_same_v<CLayout, Col>)
{
if(MRaw_ % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
return false;
}
}
else
{
return false;
}
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
......@@ -630,29 +719,12 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
const auto KRaw = arg.raw_lengths_m_n_k_o_[2];
const auto Gemm1NRaw = arg.raw_lengths_m_n_k_o_[3];
// Check scalar per vector requirement
const auto a_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
const auto b_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
const auto b1_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
const auto c_extent_lowest =
is_same_v<tensor_layout::gemm::RowMajor, CLayout> ? Gemm1NRaw : MRaw;
if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
{
return false;
}
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
arg.block_2_ctile_map_) and
IsSupported(MRaw, NRaw, KRaw, Gemm1NRaw);
}
// polymorphic
......@@ -771,6 +843,268 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return str.str();
}
#endif
template <class ADesc, class BDesc, class B1Desc, class CDesc>
struct Descriptor
{
template <class AGridDescriptor>
static constexpr auto MakeAGridDescriptor_AK0_M_AK1(const AGridDescriptor& a_grid_desc)
{
const auto a_grid_desc_m_k = DeviceOp::matrix_padder.PadADescriptor_M_K(a_grid_desc);
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
const auto AK0 = K / AK1;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
template <class BGridDescriptor>
static constexpr auto MakeBGridDescriptor_BK0_N_BK1(const BGridDescriptor& b_grid_desc)
{
const auto b_grid_desc_n_k = DeviceOp::matrix_padder.PadBDescriptor_N_K(b_grid_desc);
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
const auto BK0 = K / BK1;
return transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
template <class B1GridDescriptor>
static constexpr auto MakeB1GridDescriptor_BK0_N_BK1(const B1GridDescriptor& b1_grid_desc)
{
const auto b1_grid_desc_n_k = DeviceOp::matrix_padder.PadB1Descriptor_N_K(b1_grid_desc);
const auto N = b1_grid_desc_n_k.GetLength(I0);
const auto K = b1_grid_desc_n_k.GetLength(I1);
const auto B1K0 = K / B1K1;
return transform_tensor_descriptor(
b1_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
template <class CGridDescriptor>
static constexpr auto MakeCGridDescriptor_M_N(const CGridDescriptor& c_grid_desc)
{
return DeviceOp::matrix_padder.PadCDescriptor_M_N(c_grid_desc);
}
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(MakeAGridDescriptor_AK0_M_AK1(ADesc{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1(BDesc{}))>;
using B1GridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(MakeB1GridDescriptor_BK0_N_BK1(B1Desc{}))>;
using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(CDesc{}))>;
// GridwiseGemm
using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
B1GridDesc_BK0_N_BK1,
CGridDesc_M_N,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
Gemm1NPerBlock,
Gemm1KPerBlock,
AK1,
BK1,
B1K1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
Gemm1NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
true,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
true,
BBlockLdsExtraN,
B1BlockTransferThreadClusterLengths_BK0_N_BK1,
B1BlockTransferThreadClusterArrangeOrder,
B1BlockTransferSrcAccessOrder,
B1BlockTransferSrcVectorDim,
B1BlockTransferSrcScalarPerVector,
B1BlockTransferDstScalarPerVector_BK1,
false,
B1BlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopSched,
matrix_padder.PadN,
MaskOutUpperTriangle>;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1;
CGridDesc_M_N c_grid_desc_m_n;
C0MatrixMask c0_matrix_mask;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_descriptor_mblock_mperblock_nblock_nperblock;
// element-wise op
AElementwiseOperation a_element_op;
BElementwiseOperation b_element_op;
B1ElementwiseOperation b1_element_op;
CElementwiseOperation c_element_op;
bool has_main_k_block_loop = true;
bool is_valid = false;
constexpr Descriptor(ADesc a,
BDesc b,
B1Desc b1,
CDesc c,
AElementwiseOperation a_element_op_,
BElementwiseOperation b_element_op_,
B1ElementwiseOperation b1_element_op_,
CElementwiseOperation c_element_op_)
: a_grid_desc_ak0_m_ak1{MakeAGridDescriptor_AK0_M_AK1(a)},
b_grid_desc_bk0_n_bk1{MakeBGridDescriptor_BK0_N_BK1(b)},
b1_grid_desc_bk0_n_bk1{MakeB1GridDescriptor_BK0_N_BK1(b1)},
c_grid_desc_m_n{MakeCGridDescriptor_M_N(c)},
block_2_ctile_map{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n)},
c_grid_descriptor_mblock_mperblock_nblock_nperblock{
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n)},
has_main_k_block_loop{GridwiseGemm::CalculateHasMainKBlockLoop(
a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2))},
c0_matrix_mask{c.GetLength(I1)},
a_element_op{a_element_op_},
b_element_op{b_element_op_},
b1_element_op{b1_element_op_},
c_element_op{c_element_op_},
is_valid{GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
b1_grid_desc_bk0_n_bk1,
c_grid_desc_m_n,
block_2_ctile_map) and
IsSupported(a_grid_desc_ak0_m_ak1.GetLength(I1),
b_grid_desc_bk0_n_bk1.GetLength(I1),
a_grid_desc_ak0_m_ak1.GetLength(I0) *
a_grid_desc_ak0_m_ak1.GetLength(I2),
b1_grid_desc_bk0_n_bk1.GetLength(I1))}
{
}
constexpr bool IsValid() const { return is_valid; }
};
template <class ADesc, class BDesc, class B1Desc, class CDesc>
static constexpr auto
make_descriptor(ADesc a,
BDesc b,
B1Desc b1,
CDesc c,
AElementwiseOperation a_element_op = AElementwiseOperation{},
BElementwiseOperation b_element_op = BElementwiseOperation{},
B1ElementwiseOperation b1_element_op = B1ElementwiseOperation{},
CElementwiseOperation c_element_op = CElementwiseOperation{})
{
return Descriptor<ADesc, BDesc, B1Desc, CDesc>(
a, b, b1, c, a_element_op, b_element_op, b1_element_op, c_element_op);
}
template <class Desc>
__device__ static void Run(const Desc& desc,
const float scale,
const ADataType* __restrict__ p_a_grid,
const ADataType* __restrict__ p_b_grid,
const ADataType* __restrict__ p_b1_grid,
CDataType* __restrict__ p_c_grid)
{
#ifndef __HIPCC_RTC__
assert(desc.is_valid);
#endif
__shared__ char p_shared_block[Desc::GridwiseGemm::GetSharedMemoryNumberOfByte()];
AccElementwiseOperation acc_element_op{scale};
if(desc.has_main_k_block_loop)
{
Desc::GridwiseGemm::template Run<true>(
p_a_grid,
p_b_grid,
p_b1_grid,
p_c_grid,
p_shared_block,
desc.a_element_op,
desc.b_element_op,
acc_element_op,
desc.b1_element_op,
desc.c_element_op,
desc.a_grid_desc_ak0_m_ak1,
desc.b_grid_desc_bk0_n_bk1,
desc.b1_grid_desc_bk0_n_bk1,
desc.c_grid_descriptor_mblock_mperblock_nblock_nperblock,
desc.block_2_ctile_map,
desc.c0_matrix_mask);
}
else
{
Desc::GridwiseGemm::template Run<false>(
p_a_grid,
p_b_grid,
p_b1_grid,
p_c_grid,
p_shared_block,
desc.a_element_op,
desc.b_element_op,
acc_element_op,
desc.b1_element_op,
desc.c_element_op,
desc.a_grid_desc_ak0_m_ak1,
desc.b_grid_desc_bk0_n_bk1,
desc.b1_grid_desc_bk0_n_bk1,
desc.c_grid_descriptor_mblock_mperblock_nblock_nperblock,
desc.block_2_ctile_map,
desc.c0_matrix_mask);
}
}
};
} // namespace device
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
View file @ 2a30cfdd
......@@ -52,8 +52,7 @@ __global__ void
#endif
kernel_batched_gemm_xdlops_v2r3(const typename DeviceOp::Argument karg)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / karg.Batch);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl_fpAintB_b_scale.hpp 0 → 100644
View file @ 2a30cfdd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_b_scale.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/flush_cache.hpp"
namespace ck {
// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
// kernel function Blockers:
// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
// two lds chunks.
// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
// buffer when we declare __shared__ inside blkgemmpipe
template <typename GridwiseGemm,
typename BatchedGemmArg,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
kernel_batched_gemm_b_scale_xdl_cshuffle_v3(BatchedGemmArg karg)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t g_idx = blockIdx.z % karg.Batch;
const index_t k_idx = blockIdx.z / karg.Batch;
const auto a_batch_offset = karg.compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const auto b_batch_offset = karg.compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const auto c_batch_offset = karg.compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
const auto b_scale_batch_offset = karg.compute_ptr_offset_of_batch.GetSacleBPtrOffset(g_idx);
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, k_idx);
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + a_batch_offset + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + b_batch_offset + splitk_batch_offset.b_k_split_offset,
karg.p_c_grid + c_batch_offset + splitk_batch_offset.c_reduce_offset,
karg.p_b_scale_grid + b_scale_batch_offset + splitk_batch_offset.scale_k_split_offset,
p_shared,
karg);
#else
ignore = karg;
#endif // end of if (defined(__gfx9__))
}
template <typename GridwiseGemm,
typename BatchedGemmArg,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
index_t MinimumOccupancy = 1,
TailNumber TailNum = TailNumber::Full>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
kernel_batched_gemm_b_scale_xdl_cshuffle_v3_2lds(BatchedGemmArg karg)
{
#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()];
__shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t g_idx = blockIdx.z % karg.Batch;
const index_t k_idx = blockIdx.z / karg.Batch;
const auto a_batch_offset = karg.compute_ptr_offset_of_batch.GetAPtrOffset(g_idx);
const auto b_batch_offset = karg.compute_ptr_offset_of_batch.GetBPtrOffset(g_idx);
const auto c_batch_offset = karg.compute_ptr_offset_of_batch.GetCPtrOffset(g_idx);
const auto b_scale_batch_offset = karg.compute_ptr_offset_of_batch.GetSacleBPtrOffset(g_idx);
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, k_idx);
GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_a_grid + a_batch_offset + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + b_batch_offset + splitk_batch_offset.b_k_split_offset,
karg.p_c_grid + c_batch_offset + splitk_batch_offset.c_reduce_offset,
karg.p_b_scale_grid + b_scale_batch_offset + splitk_batch_offset.scale_k_split_offset,
p_shared_0,
p_shared_1,
karg);
#else
ignore = karg;
#endif // end of if (defined(__gfx9__))
}
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename BScaleDataType,
typename CDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization GemmSpec,
index_t BlockSize,
index_t ScaleBlockN, // scale block for N
index_t ScaleBlockK, // scale block for K
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1,
index_t BK1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA,
bool PermuteA = false,
bool PermuteB = false>
struct DeviceBatchedGemm_Xdl_CShuffleV3_BScale
: public DeviceBatchedGemmV2BScale<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
BScaleDataType,
CDataType,
ScaleBlockN,
ScaleBlockK,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
{
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_xdl_cshuffle_v3<
ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
GemmAccDataType,
CShuffleDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
BlockSize,
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,
CShuffleBlockTransferScalarPerVector_NPerBlock,
BlkGemmPipeSched,
BlkGemmPipelineVer,
ComputeTypeA,
ComputeTypeB,
PermuteA,
PermuteB>;
static constexpr index_t BPackedSize = []() {
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
return 2;
else
return 1;
}();
struct ComputePtrOffsetOfStridedBatch
{
ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideC,
index_t BatchStrideScaleB)
: BatchStrideA_(BatchStrideA),
BatchStrideB_(BatchStrideB),
BatchStrideC_(BatchStrideC),
BatchStrideScaleB_(BatchStrideScaleB)
{
}
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideA_);
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB_) / BPackedSize;
}
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
__host__ __device__ constexpr long_index_t GetSacleBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideScaleB_);
}
private:
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideC_;
index_t BatchStrideScaleB_;
};
struct Argument : public GridwiseGemm::Argument
{
index_t Batch;
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch;
Argument(const ADataType* p_a_grid_,
const BDataType* p_b_grid_,
CDataType* p_c_grid_,
index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t StrideScaleB_,
index_t BatchStrideA_,
index_t BatchStrideB_,
index_t BatchStrideC_,
index_t BatchStrideScaleB_,
const BScaleDataType* p_b_scale_grid_,
index_t Batch_,
index_t KBatch_,
AElementwiseOperation a_element_op_,
BElementwiseOperation b_element_op_,
CElementwiseOperation c_element_op_)
: GridwiseGemm::Argument(p_a_grid_,
p_b_grid_,
p_c_grid_,
M_,
N_,
K_,
StrideA_,
StrideB_,
StrideC_,
StrideScaleB_,
p_b_scale_grid_,
KBatch_, // KBatch
a_element_op_,
b_element_op_,
c_element_op_),
Batch(Batch_),
compute_ptr_offset_of_batch(
BatchStrideA_, BatchStrideB_, BatchStrideC_, BatchStrideScaleB_)
{
}
};
// Invoker
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(stream_config.log_level_ > 0)
{
arg.Print();
}
if(!GridwiseGemm::CheckValidity(arg))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
index_t gdx, gdy, gdz;
std::tie(gdx, gdy, gdz) =
GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.Batch * arg.KBatch);
float ave_time = 0;
index_t k_grain = arg.KBatch * KPerBlock;
index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
const auto Run = [&](const auto& kernel) {
if(stream_config.flush_cache)
{
Argument arg_ = arg;
const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
auto size_a_buffer =
a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
auto size_b_buffer =
b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
ck::utility::RotatingMemWrapper<Argument> rotating_mem(
arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck::utility::flush_icache();
// rotating mem
rotating_mem.Next();
// clear c mem
if(arg_.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
0,
arg_.M * arg_.N * sizeof(CDataType),
stream_config.stream_id_));
};
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
stream_config,
run_flush_cache,
kernel,
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg_);
}
else
{
if(arg.KBatch > 1)
hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
0,
arg.M * arg.N * sizeof(CDataType),
stream_config.stream_id_));
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
}
};
constexpr index_t minimum_occupancy =
BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave
? (BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 &&
MPerBlock * NPerBlock * KPerBlock * sizeof(ADataType) <= 128 * 128 * 64 * 2)
? 2
: 1
: 2;
if(has_main_k_block_loop)
{
// Tail number always full
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
if(arg.KBatch > 1)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy>;
Run(kernel);
}
else
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
Run(kernel);
}
}
// Tail number could be One to Seven
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::One>;
Run(kernel);
}
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Full)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Full>;
Run(kernel);
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Two>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Three)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Three>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Four)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Four>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Five)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Five>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Six>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Seven)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Seven>;
Run(kernel);
}
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::One>;
Run(kernel);
}
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Full)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Full>;
Run(kernel);
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Two>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Three)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Three>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Four)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Four>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Five)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Five>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Six>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Seven)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Seven>;
Run(kernel);
}
}
}
}
// Tail number could be Odd or Even
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3_2lds<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3_2lds<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3_2lds<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3_2lds<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
else
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
}
else
{
// Tail number always 1
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
if(arg.KBatch > 1)
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
false,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy>;
Run(kernel);
}
else
{
const auto kernel = kernel_batched_gemm_b_scale_xdl_cshuffle_v3<
GridwiseGemm,
Argument,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
Run(kernel);
}
}
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
{
return false;
}
if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
{
return false;
}
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding ||
GemmSpec == GemmSpecialization::KPadding))
{
return false;
}
return GridwiseGemm::CheckValidity(arg);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
index_t GetKPerBlock() override { return KPerBlock; }
bool GetPermuteB() override { return PermuteB; }
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t StrideScaleB,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideC,
index_t BatchStrideScaleB,
const BScaleDataType* p_b_scale,
index_t Batch,
index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{p_a,
p_b,
p_c,
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideScaleB,
BatchStrideA,
BatchStrideB,
BatchStrideC,
BatchStrideScaleB,
p_b_scale,
Batch,
KBatch,
a_element_op,
b_element_op,
c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t StrideScaleB,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideC,
index_t BatchStrideScaleB,
const void* p_b_scale,
index_t Batch,
index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideScaleB,
BatchStrideA,
BatchStrideB,
BatchStrideC,
BatchStrideScaleB,
static_cast<const BScaleDataType*>(p_b_scale),
Batch,
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
{BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
{BlockGemmPipelineScheduler::Interwave, "Interwave"}};
std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
{BlockGemmPipelineVersion::v1, "v1"},
{BlockGemmPipelineVersion::v2, "v2"},
{BlockGemmPipelineVersion::v3, "v3"},
{BlockGemmPipelineVersion::v4, "v4"},
{BlockGemmPipelineVersion::v5, "v5"}};
// clang-format off
str << "DeviceGemmXdlUniversal"
<< "<"
<< getGemmSpecializationString(GemmSpec) << ", "
<< std::string(ALayout::name)[0]
<< std::string(BLayout::name)[0]
<< std::string(CLayout::name)[0]
<< ">"
<< " BlkSize: "
<< BlockSize << ", "
<< "BlkTile: "
<< MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
<< "WaveTile: "
<< MPerXDL<<"x"<<NPerXDL << ", "
<< "WaveMap: "
<< MXdlPerWave<<"x" << NXdlPerWave<<", "
<< "VmemReadVec: "
<< ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
<< "BlkGemmPipelineScheduler: "
<< BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
<< "BlkGemmPipelineVersion: "
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< "BlkGemmPipelinePrefetchStages: "
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
View file @ 2a30cfdd
......@@ -3,6 +3,7 @@
#pragma once
#include "ck/library/utility/numeric.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp"
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -55,8 +55,7 @@ __global__ void
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_as_grid,
......
include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
View file @ 2a30cfdd
......@@ -55,8 +55,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / num_batches);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
......@@ -97,7 +96,7 @@ __global__ void
ignore = b_element_op;
ignore = c_element_op;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
#endif // end of if (defined(__gfx9__))
}
// specialization for #D conv: in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
View file @ 2a30cfdd
......@@ -50,9 +50,8 @@ __global__ void
const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__) || \
defined(__gfx12__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx9__) || \
defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -63,8 +63,7 @@ __global__ void
const Block2ETileMap block_2_etile_map,
index_t NRaw)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemmWelford::GetSharedMemoryNumberOfByte()];
GridwiseGemmWelford::template Run<HasMainKBlockLoop>(
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -60,8 +60,7 @@ __global__ void
const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
View file @ 2a30cfdd
......@@ -54,8 +54,7 @@ __global__ void
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp 100755 → 100644
View file @ 2a30cfdd
......@@ -469,7 +469,11 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout
{
return false;
}
if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> &&
arg.Streamk_sel > 0)
{
return false;
}
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding ||
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp
View file @ 2a30cfdd
......@@ -64,7 +64,9 @@ template <typename ALayout,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA>
typename ComputeTypeB = ComputeTypeA,
bool PermuteA = false,
bool PermuteB = false>
struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
BLayout,
CLayout,
......@@ -122,7 +124,9 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
BlkGemmPipeSched,
BlkGemmPipelineVer,
ComputeTypeA,
ComputeTypeB>;
ComputeTypeB,
PermuteA,
PermuteB>;
using Argument = typename GridwiseGemm::Argument;
......@@ -134,6 +138,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
if(stream_config.log_level_ > 0)
{
arg.Print();
GridwiseGemm::BlockwiseGemmPipe::HotLoopInstList::Print();
}
if(!GridwiseGemm::CheckValidity(arg))
......@@ -633,6 +638,11 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
index_t GetKPerBlock() override { return KPerBlock; }
bool GetPermuteA() override { return PermuteA; }
bool GetPermuteB() override { return PermuteB; }
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
......@@ -724,11 +734,14 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
<< "BlkGemmPipelineVersion: "
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< "BlkGemmPipelinePrefetchStages: "
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages << ", "
<< "Kpack: "
<< GridwiseGemm::BlockwiseGemmPipe::AMmaKStride;
// clang-format on
return str.str();
}
REGISTER_EXTRA_PRINTING_METHODS
};
} // namespace device
......
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