Skip to content

GitLab

Commit b66c4cc7 authored by Adam Osewski's avatar Adam Osewski
Browse files

Use GridwiseGemmPipeline

parent 3948d09b
Hide whitespace changes
Inline Side-by-side
Showing with 42 additions and 29 deletions
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ b66c4cc7
......@@ -73,6 +73,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
// TODO: should be exposed as Tparams.
static constexpr index_t NumGemmKPrefetchStage = 1;
static constexpr LoopScheduler LoopSched = make_default_loop_scheduler();
static constexpr PipelineVersion PipelineVer = PipelineVersion::v2;
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
......@@ -85,6 +90,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
NumGemmKPrefetchStage,
MPerBlock,
NPerBlock,
K0PerBlock,
......@@ -112,7 +118,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CBlockTransferScalarPerVector_NWaveNPerXDL,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
LoopSched,
PipelineVer>;
using Argument = typename GridwiseGemm::Argument;
using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
View file @ b66c4cc7
......@@ -85,7 +85,7 @@ template <typename ALayout,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
GemmSpecialization GemmSpec,
ck::index_t NumPrefetch,
ck::index_t NumGemmKPrefetchStage,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
......@@ -152,6 +152,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
BElementwiseOperation,
CDEElementwiseOperation,
GemmSpec,
NumGemmKPrefetchStage,
MPerBlock,
NPerBlock,
K0PerBlock,
......@@ -179,7 +180,9 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CDEBlockTransferScalarPerVector_NPerBlock,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
LoopSched,
PipelineVersion::v2>;
using CGridDesc_M_N = typename GridwiseGemm::CGridDesc_M_N;
using Block2ETileMapKSplit =
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ b66c4cc7
......@@ -55,6 +55,7 @@ template <index_t BlockSize,
typename BElementwiseOperation,
typename CElementwiseOperation,
tensor_operation::device::GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
index_t MPerBlock,
index_t NPerBlock,
index_t K0PerBlock,
......@@ -82,7 +83,9 @@ template <index_t BlockSize,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>
typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler(),
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
{
static constexpr auto I0 = Number<0>{};
......@@ -101,6 +104,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
struct Argument : public ck::tensor_operation::device::BaseArgument
{
const FloatAB* p_a_grid;
......@@ -497,6 +503,18 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
}
}
const auto num_k_loop = karg.K0 / K0PerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
#if DEBUG_LOG
std::cout << "The number of k loops (" << num_k_loop
<< ") value is not supported by GridwiseGemm Pipeline."
<< " K0: " << karg.K0 << ", K0PerBlock: " << K0PerBlock << " " << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__ << std::endl;
#endif // DEBUG_LOG
return false;
}
return true;
}
......@@ -509,9 +527,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{
const bool has_main_k0_block_loop = K0 > K0PerBlock;
return has_main_k0_block_loop;
const index_t num_loop = K0 / K0PerBlock;
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
template <typename CGridDesc>
......@@ -750,20 +767,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
#if 1
auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatAcc,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
K1>{};
#else
auto blockwise_gemm = BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
FloatAB,
FloatAcc,
......@@ -773,9 +778,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
NPerXDL,
MRepeat,
NRepeat,
K1>{};
#endif
K1,
LoopSched>();
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
......@@ -831,7 +835,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
k0_block_data_begin += K0PerBlock;
} while(k0_block_data_begin < (K0 - K0PerBlock));
} while(k0_block_data_begin < (karg.K0 - K0PerBlock));
}
// tail
......@@ -841,14 +845,12 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
#else
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_Selector<PipelineVersion::v2, 1, LoopScheduler::Default>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
(K0PerBlock * K1));
const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
a_b_k0_m_k1_block_desc,
a_blockwise_copy,
......
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