Skip to content

GitLab

"git@developer.sourcefind.cn:zhaoyu6/sglang.git" did not exist on "4fece12be98292ded0ddf05d8068d5647fcebf50"
Unverified Commit bde8e668 authored by Rostyslav Geyyer's avatar Rostyslav Geyyer Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-674

parents b95deabb f53ede26
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 913 additions and 433 deletions
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
View file @ bde8e668
......@@ -273,7 +273,10 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
cde_element_op_{cde_element_op},
MRaw_{M},
NRaw_{N},
KRaw_{K}
{
a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
......@@ -335,6 +338,11 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
// for checking vector load/store
index_t MRaw_;
index_t NRaw_;
index_t KRaw_;
};
// Invoker
......@@ -488,6 +496,85 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
{
return false;
}
// 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> && ABlockTransferSrcVectorDim == 2)
{
if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
{
// FIXME: not rigorous
if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector laod of B
if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
{
if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
{
// FIXME: not rigorous
if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector load of Ds
// only support RowMajor for now
bool all_valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
if constexpr(!is_same_v<DLayout, Row>)
{
all_valid = false;
}
});
if(!all_valid)
{
return false;
}
// check vector store of E
// only support RowMajor for now
if constexpr(is_same_v<ELayout, Row>)
{
if(arg.NRaw_ % CDEShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
return false;
}
}
else
{
return false;
}
}
return GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
View file @ bde8e668
......@@ -51,7 +51,8 @@ __global__ void
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......@@ -490,7 +491,8 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
View file @ bde8e668
......@@ -239,7 +239,10 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op}
c_element_op_{c_element_op},
MRaw_{M},
NRaw_{N},
KRaw_{K}
{
a_grid_desc_k0_m_k1_ =
DeviceGemmWmma_CShuffle::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
......@@ -276,6 +279,10 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
// for checking vector load/store
index_t MRaw_;
index_t NRaw_;
index_t KRaw_;
};
// Invoker
......@@ -417,6 +424,68 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
return false;
}
// 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> && ABlockTransferSrcVectorDim == 2)
{
if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
{
// FIXME: not rigorous
if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector laod of B
if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
{
if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
{
// FIXME: not rigorous
if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector store of C
// only support RowMajor for now
if constexpr(is_same_v<CLayout, Row>)
{
if(arg.NRaw_ % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
return false;
}
}
else
{
return false;
}
}
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
View file @ bde8e668
......@@ -428,7 +428,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
return false;
}
}
else if(ck::get_device_name() == "gfx90a")
else if(ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
View file @ bde8e668
......@@ -574,7 +574,8 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
View file @ bde8e668
......@@ -648,7 +648,8 @@ struct DeviceGemmLayerNorm_Xdl_CShuffle : public BaseOperator
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
View file @ bde8e668
......@@ -37,7 +37,8 @@ __global__ void
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
......@@ -703,7 +704,8 @@ struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940"))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
View file @ bde8e668
......@@ -130,7 +130,8 @@ __global__ void
const Block2ETileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp
View file @ bde8e668
......@@ -78,7 +78,8 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
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_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
View file @ bde8e668
......@@ -155,7 +155,8 @@ __global__ void
const Block2ETileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
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);
......@@ -810,7 +811,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
return false;
}
}
else if(get_device_name() == "gfx90a")
else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
View file @ bde8e668
......@@ -135,7 +135,8 @@ __global__ void
const Block2ETileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -684,7 +685,7 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
return false;
}
}
else if(get_device_name() == "gfx90a")
else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
View file @ bde8e668
......@@ -38,7 +38,8 @@ __global__ void
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation c_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ bde8e668
......@@ -587,4 +587,52 @@ struct OffsettedBlockToCTileMap
index_t block_start_;
};
/**
* @brief Simple tile mapping which creates 3D grid of block of threads.
*
* @paragraph Description
* This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
* blocks. The first 2D are regular 2D tiles created by division of output GEMM
* dimenions by corresponding tile size. The third dimension (Z) is a k-split dimension,
* which denotes the number of blocks we use to divide work on GEMM K dimension onto.
*
* @tparam MPerBlock Output block tile size in M dimension.
* @tparam NPerBlock Output block tile size in N dimension.
*/
template <index_t MPerBlock, index_t NPerBlock>
struct BlockToCTileMap_3DGrid_KSplit
{
__host__ __device__ BlockToCTileMap_3DGrid_KSplit() = default;
__host__ __device__ constexpr auto
CalculateGridSize(index_t M, index_t N, index_t k_split) const
{
// Create 3D grid
const auto M0 = math::integer_divide_ceil(M, MPerBlock);
const auto N0 = math::integer_divide_ceil(N, NPerBlock);
return std::make_tuple(N0, M0, k_split);
}
template <typename TopIdx>
__device__ constexpr auto CalculateBottomIndex(const TopIdx&) const
{
return make_tuple(blockIdx.z, blockIdx.y, blockIdx.x);
}
template <typename CTileIdx, typename CTileDim>
__host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
const CTileDim& /* c_tile_dim */) const
{
return true; // always valid provided that user gets grid size from CalculateGridSize()
}
template <typename CGridDesc_M_N>
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
{
return true;
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
View file @ bde8e668
......@@ -66,7 +66,8 @@ __global__ void
const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
View file @ bde8e668
......@@ -54,7 +54,8 @@ __global__ void
const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
View file @ bde8e668
......@@ -44,7 +44,8 @@ __global__ void
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
View file @ bde8e668
......@@ -57,7 +57,8 @@ __global__ void
const C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
// TODO ANT: separate into MMA + Epilogue
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
View file @ bde8e668
......@@ -165,7 +165,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
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