Skip to content

GitLab

Unverified Commit b164b0ef authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into grouped_gemm_multi_abd_fixed_nk_example

parents e786932a 3696fe1c
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2474 additions and 243 deletions
example/CMakeLists.txt
View file @ b164b0ef
......@@ -62,6 +62,12 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
set(result ${result} PARENT_SCOPE)
endfunction(add_example_executable EXAMPLE_NAME)
function(add_example_dependencies EXAMPLE_NAME FILE_NAME)
if(result EQUAL 0)
add_dependencies(${EXAMPLE_NAME} ${FILE_NAME})
endif()
endfunction(add_example_dependencies EXAMPLE_NAME)
function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
set(result 1)
......
include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp 0 → 100644
View file @ b164b0ef
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// GEMM:
// input : A0[M0, M1, ... K0, K1, ...], ...
// input : B0[N0, N1, ... K0, K1, ...], ...
// input : D0[M0, M1, ... N0, N1, ...], D1[M0, M1, ... N0, N1, ...], ...
// output : E[M0, M1, ... N0, N1, ...]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// D0, D1, ... and E have the same layout
template <index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename AsDataType,
typename BsDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceContractionMultipleABD : public BaseOperator
{
static constexpr index_t NumATensor = AsDataType::Size();
static constexpr index_t NumBTensor = BsDataType::Size();
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
std::array<const void*, NumBTensor> p_bs,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_lengths,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_strides,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_lengths,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_strides,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_strides,
const std::vector<index_t>& e_ms_ns_length,
const std::vector<index_t>& e_ms_ns_stride,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
View file @ b164b0ef
......@@ -20,7 +20,8 @@ template <typename ALayout,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
typename ComputeType = CDataType>
struct DeviceGemmSplitK : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
......@@ -48,7 +49,8 @@ template <typename ALayout,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
typename ComputeType = CDataType>
using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
BLayout,
CLayout,
......@@ -57,7 +59,8 @@ using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>>;
CElementwiseOperation,
ComputeType>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_normalization.hpp
View file @ b164b0ef
......@@ -14,8 +14,8 @@ namespace device {
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename ComputeDataType,
typename YDataType,
typename SaveMeanInvStdDataType,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
......@@ -27,6 +27,8 @@ struct DeviceNormalization : public BaseOperator
const std::vector<index_t> gammaStrides,
const std::vector<index_t> betaStrides,
const std::vector<index_t> yStrides,
const std::vector<index_t> saveMeanStrides,
const std::vector<index_t> saveInvStdStrides,
const std::vector<index_t> reduceDims,
double epsilon,
const void* p_x,
......@@ -43,16 +45,16 @@ struct DeviceNormalization : public BaseOperator
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename ComputeDataType,
typename YDataType,
typename SaveMeanInvStdDataType,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementwiseOperation,
Rank,
NumReduceDim>>;
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ b164b0ef
......@@ -69,7 +69,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
CElementwiseOperation,
ComputeType>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -126,7 +127,50 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
PipelineVer,
ComputeType>;
using Argument = typename GridwiseGemm::Argument;
struct Argument : public GridwiseGemm::Argument
{
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 MPadded_,
index_t NPadded_,
index_t KPadded_,
index_t K0_,
index_t k_batch_,
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_,
MPadded_,
NPadded_,
KPadded_,
K0_,
k_batch_),
a_element_op(a_element_op_),
b_element_op(b_element_op_),
c_element_op(c_element_op_)
{
}
AElementwiseOperation a_element_op;
BElementwiseOperation b_element_op;
CElementwiseOperation c_element_op;
};
using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
// Invoker
......@@ -167,8 +211,17 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
karg.M * karg.N * sizeof(CDataType),
stream_config.stream_id_));
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg, b2c_map);
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
static_cast<typename GridwiseGemm::Argument>(karg),
b2c_map,
karg.a_element_op,
karg.b_element_op,
karg.c_element_op);
};
if(has_main_k0_block_loop)
......@@ -179,7 +232,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -189,7 +245,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -202,7 +261,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -212,7 +274,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
false,
InMemoryDataOperationEnum::AtomicAdd,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -260,12 +325,12 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t KBatch)
{
return Argument{p_a,
return Argument(p_a,
p_b,
p_c,
M,
......@@ -278,7 +343,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
GridwiseGemm::CalculateNPadded(N),
GridwiseGemm::CalculateKPadded(K, KBatch),
GridwiseGemm::CalculateK0(K, KBatch),
KBatch};
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -293,9 +361,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t KBatch = 1) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
......@@ -311,7 +379,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
GridwiseGemm::CalculateNPadded(N),
GridwiseGemm::CalculateKPadded(K, KBatch),
GridwiseGemm::CalculateK0(K, KBatch),
KBatch);
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
View file @ b164b0ef
......@@ -565,7 +565,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
auto launch_kernel = [&](auto has_main_k_block_loop) {
constexpr bool has_main_loop = has_main_k_block_loop.value;
const auto kernel = kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle<
const auto kernel = kernel_grouped_conv_multiple_d_wmma_cshuffle<
GridwiseGemm,
ADataType,
BDataType,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
View file @ b164b0ef
......@@ -12,6 +12,7 @@
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
......@@ -22,32 +23,6 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace {
struct ComputePtrOffsetOfStridedBatch
{
__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_);
}
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideC_;
};
} // namespace
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
......@@ -952,7 +927,7 @@ struct DeviceGroupedConvBwdWeight_Dl : public DeviceGroupedConvBwdWeight<NDimSpa
Block2CTileMap block_2_ctile_map_;
// for computing batch offset
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
ComputePtrOffsetOfStridedBatch<I0> compute_ptr_offset_of_batch_;
// element-wise op
OutElementwiseOperation a_element_op_;
......@@ -1024,7 +999,7 @@ struct DeviceGroupedConvBwdWeight_Dl : public DeviceGroupedConvBwdWeight<NDimSpa
remove_reference_t<DeviceOp::BGridDesc_B_K0_N0_N1_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11>,
remove_reference_t<DeviceOp::Block2CTileMap>,
ComputePtrOffsetOfStridedBatch,
ComputePtrOffsetOfStridedBatch<I0>,
has_main_loop,
has_double_loop>;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
View file @ b164b0ef
......@@ -14,6 +14,7 @@
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -21,32 +22,6 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace {
struct ComputePtrOffsetOfStridedBatch
{
__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_);
}
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideC_;
};
} // namespace
template <typename GridwiseGemm,
typename FloatA,
typename FloatB,
......@@ -1222,7 +1197,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
Block2CTileMap block_2_ctile_map_;
// for computing batch offset
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
ComputePtrOffsetOfStridedBatch<I0> compute_ptr_offset_of_batch_;
index_t M01_;
index_t N01_;
......@@ -1301,7 +1276,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<DeviceOp::Block2CTileMap>,
ComputePtrOffsetOfStridedBatch,
ComputePtrOffsetOfStridedBatch<I0>,
has_main_loop>;
return launch_and_time_kernel(stream_config,
......@@ -1348,6 +1323,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
{
return false;
}
if constexpr(NDimSpatial == 1)
{
if constexpr(!is_GNWK_GKXC_GNWC)
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
View file @ b164b0ef
......@@ -471,7 +471,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
auto launch_kernel = [&](auto has_main_k_block_loop) {
constexpr bool has_main_loop = has_main_k_block_loop.value;
const auto kernel = kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle<
const auto kernel = kernel_grouped_conv_multiple_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp
View file @ b164b0ef
......@@ -43,7 +43,13 @@ struct ComputePtrOffsetOfStridedBatch
return ds_offset;
}
__host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
[[maybe_unused]] __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideE_);
}
// alias for kernels without multiple D
[[maybe_unused]] __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideE_);
}
......@@ -52,6 +58,7 @@ struct ComputePtrOffsetOfStridedBatch
index_t BatchStrideB_;
Array<ck::index_t, NumDTensor> BatchStrideDs_;
index_t BatchStrideE_;
index_t& BatchStrideC_ = BatchStrideE_; // alias for kernels without multiple D
};
} // namespace device
......
include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
View file @ b164b0ef
......@@ -28,6 +28,7 @@ template <typename XDataType,
typename BetaDataType,
typename ComputeDataType,
typename YDataType,
typename SaveMeanInvStdDataType,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim,
......@@ -43,12 +44,13 @@ template <typename XDataType,
index_t BetaSrcVectorDim,
index_t BetaSrcVectorSize,
index_t YDstVectorSize,
index_t SaveMeanInvStdDstVectorSize,
bool UseWelford = true>
struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementwiseOperation,
Rank,
NumReduceDim>
......@@ -64,18 +66,24 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
(BetaSrcVectorDim == 1 && KThreadSliceSize % BetaSrcVectorSize == 0)),
"Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
static_assert(MThreadSliceSize % SaveMeanInvStdDstVectorSize == 0,
"Invalid thread slice sizes and/or save mean and inverse std vector sizes "
"configuration, please check!");
using PassThrough = tensor_operation::element_wise::PassThrough;
static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr bool reduceAllDim = (NumInvariantDim == 0);
static_assert(!reduceAllDim); // TODO
static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
const std::vector<index_t>& inStrides,
int numBlockTileIteration)
{
constexpr index_t NumInvariantDim = Rank - NumReduceDim;
static constexpr index_t numSrcDim = Rank;
static constexpr bool reduceAllDim = (NumInvariantDim == 0);
const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
......@@ -133,7 +141,37 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
return (in_grid_desc_m_k_padded);
};
static auto MakeSaveMeanInvStdDescriptor_M(const std::vector<index_t>& lengths,
const std::vector<index_t>& strides)
{
using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
const auto tupleSrcLengths = make_tuple_from_array_and_index_seq(lengths, InvariantDims{});
const auto tupleSrcStrides = make_tuple_from_array_and_index_seq(strides, InvariantDims{});
const auto desc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
const auto grid_desc_m =
transform_tensor_descriptor(desc,
make_tuple(make_merge_transform(tupleSrcLengths)),
make_tuple(InvariantDims{}),
make_tuple(Sequence<0>{}));
const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
const auto pad_M =
math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
auto grid_desc_m_padded = transform_tensor_descriptor(
grid_desc_m,
make_tuple(make_right_pad_transform(invariantLength, pad_M)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
return grid_desc_m_padded;
}
using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1));
using GridDesc_M = decltype(MakeSaveMeanInvStdDescriptor_M({1}, {1}));
struct Argument : public BaseArgument
{
......@@ -142,17 +180,23 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
const std::vector<index_t> gammaStrides,
const std::vector<index_t> betaStrides,
const std::vector<index_t> yStrides,
const std::vector<index_t> saveMeanStrides,
const std::vector<index_t> saveInvStdStrides,
const std::vector<index_t> reduceDims,
YElementwiseOperation y_elementwise_op,
double epsilon,
const XDataType* p_x,
const GammaDataType* p_gamma,
const BetaDataType* p_beta,
YDataType* p_y)
YDataType* p_y,
SaveMeanInvStdDataType* p_saveMean,
SaveMeanInvStdDataType* p_saveInvStd)
: p_x_(p_x),
p_gamma_(p_gamma),
p_beta_(p_beta),
p_y_(p_y),
p_saveMean_(p_saveMean),
p_saveInvStd_(p_saveInvStd),
y_elementwise_op_(y_elementwise_op)
{
epsilon_ = static_cast<ComputeDataType>(epsilon);
......@@ -162,16 +206,14 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
yStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
betaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
saveMeanStrides_ = saveMeanStrides;
saveInvStdStrides_ = saveInvStdStrides;
long_index_t invariant_length;
long_index_t reduce_length;
std::tie(invariant_length, reduce_length) =
get_2d_lengths<Rank, NumReduceDim>(Lengths_);
std::tie(MRaw_, KRaw_) = get_2d_lengths<Rank, NumReduceDim>(Lengths_);
numBlockTileIteration_ = math::integer_divide_ceil(reduce_length, K_BlockTileSize);
numBlockTileIteration_ = math::integer_divide_ceil(KRaw_, K_BlockTileSize);
gridSize_ = math::integer_divide_ceil(invariant_length, M_BlockTileSize);
gridSize_ = math::integer_divide_ceil(MRaw_, M_BlockTileSize);
x_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, xStrides_, numBlockTileIteration_);
gamma_grid_desc_m_k_ =
......@@ -179,9 +221,16 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
beta_grid_desc_m_k_ =
MakeSrc2dDescriptor(Lengths_, betaStrides_, numBlockTileIteration_);
y_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, yStrides_, numBlockTileIteration_);
save_mean_grid_desc_m_ = MakeSaveMeanInvStdDescriptor_M(Lengths_, saveMeanStrides);
save_inv_std_grid_desc_m_ = MakeSaveMeanInvStdDescriptor_M(Lengths_, saveInvStdStrides);
isSweeponce_ =
x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
if constexpr(NumInvariantDim == 0)
invariant_lowest_length_ = 1;
else
invariant_lowest_length_ = Lengths_[NumInvariantDim - 1];
}
ComputeDataType epsilon_;
......@@ -190,12 +239,16 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
const GammaDataType* p_gamma_;
const BetaDataType* p_beta_;
YDataType* p_y_;
SaveMeanInvStdDataType* p_saveMean_;
SaveMeanInvStdDataType* p_saveInvStd_;
std::vector<index_t> Lengths_;
std::vector<index_t> xStrides_;
std::vector<index_t> gammaStrides_;
std::vector<index_t> betaStrides_;
std::vector<index_t> yStrides_;
std::vector<index_t> saveMeanStrides_;
std::vector<index_t> saveInvStdStrides_;
YElementwiseOperation y_elementwise_op_;
......@@ -206,7 +259,14 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
GridDesc_M_K gamma_grid_desc_m_k_;
GridDesc_M_K beta_grid_desc_m_k_;
GridDesc_M_K y_grid_desc_m_k_;
GridDesc_M save_mean_grid_desc_m_;
GridDesc_M save_inv_std_grid_desc_m_;
bool isSweeponce_;
index_t MRaw_; // invarient length
index_t KRaw_; // reduce length
index_t invariant_lowest_length_;
};
struct Invoker : public BaseInvoker
......@@ -217,9 +277,11 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
YElementwiseOperation,
GridDesc_M_K,
GridDesc_M,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
......@@ -233,6 +295,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
BetaSrcVectorSize,
XYSrcVectorDim,
YDstVectorSize,
SaveMeanInvStdDstVectorSize,
UseWelford>(arg.isSweeponce_);
float avg_time = 0;
......@@ -245,12 +308,16 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
arg.gamma_grid_desc_m_k_,
arg.beta_grid_desc_m_k_,
arg.y_grid_desc_m_k_,
arg.save_mean_grid_desc_m_,
arg.save_inv_std_grid_desc_m_,
arg.numBlockTileIteration_,
arg.epsilon_,
arg.p_x_,
arg.p_gamma_,
arg.p_beta_,
arg.p_y_,
arg.p_saveMean_,
arg.p_saveInvStd_,
arg.y_elementwise_op_);
return (avg_time);
......@@ -267,8 +334,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
{
const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
constexpr index_t NumInvariantDim = Rank - NumReduceDim;
if constexpr(XYSrcVectorDim == 0)
{
if constexpr(NumInvariantDim == 0)
......@@ -277,13 +342,15 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
}
else
{
printf("!!!! %d\n", p_arg_->invariant_lowest_length_);
if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
return false;
if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
if(p_arg_->invariant_lowest_length_ % XSrcVectorSize != 0)
return false;
if(p_arg_->invariant_lowest_length % YDstVectorSize != 0)
if(p_arg_->invariant_lowest_length_ % YDstVectorSize != 0)
return false;
};
}
......@@ -325,7 +392,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
return (false);
if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
if(p_arg_->invariant_lowest_length_ % BetaSrcVectorSize != 0)
return (false);
}
else // if fastest dim is reduced
......@@ -337,6 +404,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
return (false);
}
if(p_arg_->invariant_lowest_length_ % SaveMeanInvStdDstVectorSize != 0)
return false;
return true;
};
......@@ -346,6 +416,8 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
const std::vector<index_t> gammaStrides,
const std::vector<index_t> betaStrides,
const std::vector<index_t> yStrides,
const std::vector<index_t> saveMeanStrides,
const std::vector<index_t> saveInvStdStrides,
const std::vector<index_t> reduceDims,
double epsilon,
const void* p_x,
......@@ -353,27 +425,30 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
const void* p_beta,
void* p_y,
void* p_saveMean,
void* p_saveInvVar,
void* p_saveInvStd,
YElementwiseOperation y_elementwise_op) override
{
// TODO
// Optional cache of the intermediate results (mean and InvVariance) during the
// forward pass could speedup in the backward
ignore = p_saveMean;
ignore = p_saveInvVar;
if(lengths.size() != Rank || xStrides.size() != Rank || gammaStrides.size() != Rank ||
betaStrides.size() != Rank || yStrides.size() != Rank ||
saveMeanStrides.size() != NumInvariantDim || saveInvStdStrides.size() != NumInvariantDim)
throw std::runtime_error("dimension is incorrect");
return std::make_unique<Argument>(lengths,
xStrides,
gammaStrides,
betaStrides,
yStrides,
saveMeanStrides,
saveInvStdStrides,
reduceDims,
y_elementwise_op,
epsilon,
static_cast<const XDataType*>(p_x),
static_cast<const GammaDataType*>(p_gamma),
static_cast<const BetaDataType*>(p_beta),
static_cast<YDataType*>(p_y));
static_cast<YDataType*>(p_y),
static_cast<SaveMeanInvStdDataType*>(p_saveMean),
static_cast<SaveMeanInvStdDataType*>(p_saveInvStd));
};
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ b164b0ef
......@@ -113,7 +113,6 @@ struct PassThrough
}
#endif
#if defined CK_ENABLE_FP8
template <>
__host__ __device__ void operator()<f8_t, f8_t>(f8_t& y, const f8_t& x) const
{
......@@ -143,9 +142,7 @@ struct PassThrough
{
y = type_convert<f8_t>(x);
}
#endif
#if defined CK_ENABLE_BF8
template <>
__host__ __device__ void operator()<bf8_t, bf8_t>(bf8_t& y, const bf8_t& x) const
{
......@@ -175,7 +172,6 @@ struct PassThrough
{
y = ck::type_convert<bf8_t>(x);
}
#endif
};
struct UnaryConvert
......@@ -204,7 +200,6 @@ struct ConvertBF16RTN
}
};
#if defined CK_ENABLE_FP8
struct ConvertF8SR
{
// convert to fp8 using stochastic rounding (SR)
......@@ -221,7 +216,6 @@ struct ConvertF8SR
y = f8_convert_sr<Y>(x);
}
};
#endif
struct Scale
{
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
View file @ b164b0ef
......@@ -36,7 +36,7 @@ __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle(
kernel_grouped_conv_multiple_d_wmma_cshuffle(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
......@@ -452,11 +452,11 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
// CheckValidity for kernels without multi D
template <typename Block2CTileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const DsGridDesc_M_N& ds_grid_desc_m_n,
const EGridDesc_M_N& e_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
......@@ -471,18 +471,6 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
const auto N = b_grid_desc_k0_n_k1.GetLength(I1);
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
bool valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
N == ds_grid_desc_m_n[i].GetLength(I1));
});
if(!valid)
{
return false;
}
if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
......@@ -517,6 +505,31 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
return true;
}
template <typename Block2CTileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const DsGridDesc_M_N& ds_grid_desc_m_n,
const EGridDesc_M_N& e_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
const auto M = a_grid_desc_k0_m_k1.GetLength(I1);
const auto N = b_grid_desc_k0_n_k1.GetLength(I1);
bool valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
N == ds_grid_desc_m_n[i].GetLength(I1));
});
if(!valid)
{
return false;
}
return CheckValidity(
a_grid_desc_k0_m_k1, b_grid_desc_k0_n_k1, e_grid_desc_m_n, block_2_ctile_map);
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / (K0PerBlock * K1);
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ b164b0ef
......@@ -22,13 +22,19 @@ namespace ck {
template <typename GridwiseGemm,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename Block2CTileMap>
typename Block2CTileMap,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
const Block2CTileMap& b2c_map)
const Block2CTileMap& b2c_map,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
......@@ -37,10 +43,13 @@ __global__ void
__shared__ uint8_t p_shared[shared_size];
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
karg, static_cast<void*>(p_shared), b2c_map);
karg, static_cast<void*>(p_shared), b2c_map, a_element_op, b_element_op, c_element_op);
#else
ignore = karg;
ignore = b2c_map;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
......@@ -577,7 +586,10 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
typename Block2CTileMap>
__device__ static void Run(const Argument& karg,
void* __restrict__ p_shared_block,
const Block2CTileMap& block_2_ctile_map)
const Block2CTileMap& block_2_ctile_map,
const AElementwiseOperation a_element_op = AElementwiseOperation{},
const BElementwiseOperation b_element_op = BElementwiseOperation{},
const CElementwiseOperation c_element_op = CElementwiseOperation{})
{
const FloatA* p_a_grid = karg.p_a_grid;
const FloatB* p_b_grid = karg.p_b_grid;
......@@ -590,9 +602,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
const AElementwiseOperation a_element_op = AElementwiseOperation{};
const BElementwiseOperation b_element_op = BElementwiseOperation{};
const CElementwiseOperation c_element_op = CElementwiseOperation{};
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
......@@ -761,8 +770,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
ComputeType,
ComputeType,
ComputeType, // ComputeType A
ComputeType, // ComputeType B
FloatAcc,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
......
include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
View file @ b164b0ef
......@@ -17,11 +17,13 @@ template <typename MeanVarDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename SaveMeanInvStdDataType,
typename ComputeDataType,
typename YElementwiseOperation,
typename MeanVarGridDesc_M_KBlock,
typename CountGridDesc_M_KBlock,
typename XYGammaBetaGridDesc_M_K,
typename SaveMeanInvStdGridDesc_M,
index_t BlockSize,
index_t MThreadClusterSize,
index_t KThreadClusterSize,
......@@ -34,7 +36,8 @@ template <typename MeanVarDataType,
index_t BetaSrcVectorDim,
index_t BetaSrcVectorSize,
index_t YDstVectorDim,
index_t YDstVectorSize>
index_t YDstVectorSize,
index_t SaveMeanInvStdDstVectorSize>
struct GridwiseNormalizationSplitK2nd
{
static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
......@@ -45,6 +48,10 @@ struct GridwiseNormalizationSplitK2nd
(YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!");
static_assert(MThreadSliceSize % SaveMeanInvStdDstVectorSize == 0,
"Invalid thread slice sizes and/or save mean and inverse std vector sizes "
"configuration, please check!");
static_assert(XSrcVectorSize == YDstVectorSize);
static_assert(XSrcVectorSize == GammaSrcVectorSize);
static_assert(XSrcVectorSize == BetaSrcVectorSize);
......@@ -69,6 +76,10 @@ struct GridwiseNormalizationSplitK2nd
static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
using ThreadBufferLengths_M = Sequence<MThreadSliceSize>;
static constexpr auto thread_buffer_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
static constexpr auto thread_buffer_desc_m_1 =
make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, I1));
......@@ -99,6 +110,8 @@ struct GridwiseNormalizationSplitK2nd
const XYGammaBetaGridDesc_M_K& gamma_grid_desc_m_k,
const XYGammaBetaGridDesc_M_K& beta_grid_desc_m_k,
const XYGammaBetaGridDesc_M_K& y_grid_desc_m_k,
const SaveMeanInvStdGridDesc_M& save_mean_grid_desc_m,
const SaveMeanInvStdGridDesc_M& save_inv_std_grid_desc_m,
index_t num_k_mean_var_count_iteration,
index_t num_k_block_tile_iteration,
index_t k_grid_size,
......@@ -110,6 +123,8 @@ struct GridwiseNormalizationSplitK2nd
const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global,
SaveMeanInvStdDataType* const __restrict__ p_save_mean_global,
SaveMeanInvStdDataType* const __restrict__ p_save_inv_std_global,
const YElementwiseOperation y_elementwise_op)
{
// Thread/Block id
......@@ -145,6 +160,12 @@ struct GridwiseNormalizationSplitK2nd
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
auto save_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_save_mean_global, save_mean_grid_desc_m.GetElementSpaceSize());
auto save_inv_std_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_save_inv_std_global, save_inv_std_grid_desc_m.GetElementSpaceSize());
// VGPR
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
in_mean_thread_buf;
......@@ -158,6 +179,7 @@ struct GridwiseNormalizationSplitK2nd
var_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
welford_count_thread_buf;
auto& inv_std_thread_buf = var_thread_buf;
auto x_thread_buf = generate_tuple(
[&](auto) {
......@@ -283,6 +305,42 @@ struct GridwiseNormalizationSplitK2nd
thread_k_cluster_id * YDstVectorSize),
y_elementwise_op);
auto threadwise_mean_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
SaveMeanInvStdDataType,
decltype(thread_buffer_desc_m),
SaveMeanInvStdGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
SaveMeanInvStdDstVectorSize, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1,
true>(
save_mean_grid_desc_m,
make_multi_index(block_m_cluster_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
auto threadwise_inv_std_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
SaveMeanInvStdDataType,
decltype(thread_buffer_desc_m),
SaveMeanInvStdGridDesc_M,
PassThroughOp,
ThreadBufferLengths_M,
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
SaveMeanInvStdDstVectorSize, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1,
true>(
save_inv_std_grid_desc_m,
make_multi_index(block_m_cluster_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize),
PassThroughOp{});
// step1: Merge mean and variance
constexpr auto mean_var_count_thread_copy_step_I0_k =
make_multi_index(I0, KThreadClusterSize);
......@@ -332,9 +390,33 @@ struct GridwiseNormalizationSplitK2nd
BlockwiseWelford::Run(
mean_thread_buf(I), var_thread_buf(I), welford_count_thread_buf(I));
inv_std_thread_buf(I) =
type_convert<ComputeDataType>(1.0f) / ck::math::sqrt(var_thread_buf(I) + epsilon);
});
// step2: normalization
// step2: save mean and inverse std for backward (optional)
if(block_k_cluster_id == 0 && thread_k_cluster_id == 0)
{
if(p_save_mean_global != nullptr)
{
threadwise_mean_store.Run(thread_buffer_desc_m,
make_tuple(I0),
mean_thread_buf,
save_mean_grid_desc_m,
save_mean_global_val_buf);
}
if(p_save_inv_std_global != nullptr)
{
threadwise_inv_std_store.Run(thread_buffer_desc_m,
make_tuple(I0),
inv_std_thread_buf,
save_inv_std_grid_desc_m,
save_inv_std_global_val_buf);
}
}
// step3: normalization
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
for(index_t k = 0; k < num_k_block_tile_iteration; ++k)
......@@ -360,7 +442,6 @@ struct GridwiseNormalizationSplitK2nd
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
......@@ -369,7 +450,7 @@ struct GridwiseNormalizationSplitK2nd
// normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor;
inv_std_thread_buf(iM);
// gamma
y_thread_buf(iK0)(Number<offset_m_k>{}) =
......
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