Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "3bd5e073cb4dc31c18440356758d57a5653b32ab"
Commit 7546eb7a authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge remote-tracking branch 'origin' into aosewski/test_ggemm_splitk

parents 8ad9c393 a1e344b1
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1126 additions and 86 deletions
example/27_layernorm/CMakeLists.txt
View file @ 7546eb7a
add_example_executable(example_layernorm_blockwise layernorm_blockwise.cpp) add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
example/27_layernorm/common.hpp 0 → 100644
View file @ 7546eb7a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
example/27_layernorm/layernorm_fp16.cpp 0 → 100644
View file @ 7546eb7a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
#include "run_layernorm_example.inc"
int main() { return run_groupnorm_example<DeviceInstance>(); }
example/27_layernorm/layernorm_splitk_fp16.cpp 0 → 100644
View file @ 7546eb7a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationSplitKImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // XScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // YScalarPerVector
#include "run_layernorm_example.inc"
int main() { return run_groupnorm_example<DeviceInstance>(); }
example/27_layernorm/layernorm_blockwise.cpp → example/27_layernorm/run_layernorm_example.inc
View file @ 7546eb7a
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #pragma once
#include <numeric>
#include <initializer_list> template <typename DeviceInstance>
#include <cstdlib> int run_groupnorm_example()
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
int main()
{ {
bool time_kernel = false; bool time_kernel = false;
...@@ -111,6 +63,10 @@ int main() ...@@ -111,6 +63,10 @@ int main()
return 1; return 1;
}; };
size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = device_instance.MakeInvokerPointer(); auto invoker_ptr = device_instance.MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel}); invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
...@@ -133,7 +89,8 @@ int main() ...@@ -133,7 +89,8 @@ int main()
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data()); y_dev.FromDevice(y.mData.data());
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results d1", 1e-3, 1e-3); pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
} }
return (pass ? 0 : 1); return (pass ? 0 : 1);
} }
example/42_groupnorm/CMakeLists.txt
View file @ 7546eb7a
add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp) add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp) add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
example/42_groupnorm/common.hpp
View file @ 7546eb7a
...@@ -12,6 +12,7 @@ ...@@ -12,6 +12,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp" #include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp" #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/fill.hpp" #include "ck/library/utility/fill.hpp"
......
example/42_groupnorm/groupnorm_splitk_fp16.cpp 0 → 100644
View file @ 7546eb7a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationSplitKImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
YElementOp,
Rank,
NumReduceDim,
256, // BlockSize
1, // ClusterM
256, // ClusterK
1, // SliceM
16, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
#include "run_groupnorm_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
example/42_groupnorm/run_groupnorm_example.inc
View file @ 7546eb7a
...@@ -73,6 +73,10 @@ int run_groupnorm_example(int argc, char* argv[]) ...@@ -73,6 +73,10 @@ int run_groupnorm_example(int argc, char* argv[])
return 1; return 1;
}; };
size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = device_instance.MakeInvokerPointer(); auto invoker_ptr = device_instance.MakeInvokerPointer();
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, true}); float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, true});
......
include/ck/ck.hpp
View file @ 7546eb7a
...@@ -175,7 +175,10 @@ ...@@ -175,7 +175,10 @@
// denorm test fix, required to work around dissue // denorm test fix, required to work around dissue
#ifndef CK_WORKAROUND_DENORM_FIX #ifndef CK_WORKAROUND_DENORM_FIX
#define CK_WORKAROUND_DENORM_FIX 0 #define CK_WORKAROUND_DENORM_FIX 0
#endif #elif
// enable only on MI200
#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
#endif // CK_WORKAROUND_DENORM_FIX
namespace ck { namespace ck {
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
View file @ 7546eb7a
...@@ -807,7 +807,7 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle ...@@ -807,7 +807,7 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
// workspace for welford intermediate mean // workspace for welford intermediate mean
workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64; workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
// workspace for welford intermediate mean // workspace for welford intermediate variance
workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64; workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
// workspace for welford intermediate count // workspace for welford intermediate count
......
include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
View file @ 7546eb7a
...@@ -10,8 +10,7 @@ ...@@ -10,8 +10,7 @@
#include "ck/tensor_operation/gpu/device/device_normalization.hpp" #include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp" #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp" #include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_normalization_selector.hpp" #include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
...@@ -20,6 +19,10 @@ namespace tensor_operation { ...@@ -20,6 +19,10 @@ namespace tensor_operation {
namespace device { namespace device {
// Y = Normalization(X, Beta, Gamma) // Y = Normalization(X, Beta, Gamma)
// M: Invarient length
// K: Reduce length (Calculate mean and variance along K dimension)
// eg. Length = [N, C, H, W], reduce dim = [C, H, W]
// Then, M = N, K = C * H * W
template <typename XDataType, template <typename XDataType,
typename GammaDataType, typename GammaDataType,
typename BetaDataType, typename BetaDataType,
...@@ -68,7 +71,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType, ...@@ -68,7 +71,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths, static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
const std::vector<index_t>& inStrides, const std::vector<index_t>& inStrides,
int blkGroupSize,
int numBlockTileIteration) int numBlockTileIteration)
{ {
constexpr index_t NumInvariantDim = Rank - NumReduceDim; constexpr index_t NumInvariantDim = Rank - NumReduceDim;
...@@ -117,10 +119,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType, ...@@ -117,10 +119,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{}); const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
const auto reduceLength = in_grid_desc_m_k.GetLength(Number<1>{}); const auto reduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
const auto inPad_M = const auto inPad_M =
math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength; math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength; const auto inPad_K = K_BlockTileSize * numBlockTileIteration - reduceLength;
auto in_grid_desc_m_k_padded = transform_tensor_descriptor( auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
in_grid_desc_m_k, in_grid_desc_m_k,
...@@ -132,7 +133,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType, ...@@ -132,7 +133,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
return (in_grid_desc_m_k_padded); return (in_grid_desc_m_k_padded);
}; };
using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1)); using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1));
struct Argument : public BaseArgument struct Argument : public BaseArgument
{ {
...@@ -162,26 +163,22 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType, ...@@ -162,26 +163,22 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims); gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
betaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims); betaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
long_index_t invariant_total_length; long_index_t invariant_length;
long_index_t reduce_total_length; long_index_t reduce_length;
std::tie(invariant_total_length, reduce_total_length) = std::tie(invariant_length, reduce_length) =
get_2d_lengths<Rank, NumReduceDim>(Lengths_); get_2d_lengths<Rank, NumReduceDim>(Lengths_);
blkGroupSize_ = 1; numBlockTileIteration_ = math::integer_divide_ceil(reduce_length, K_BlockTileSize);
numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) / gridSize_ = math::integer_divide_ceil(invariant_length, M_BlockTileSize);
M_BlockTileSize * blkGroupSize_;
x_grid_desc_m_k_ = x_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, xStrides_, numBlockTileIteration_);
MakeSrc2dDescriptor(Lengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
gamma_grid_desc_m_k_ = gamma_grid_desc_m_k_ =
MakeSrc2dDescriptor(Lengths_, gammaStrides_, blkGroupSize_, numBlockTileIteration_); MakeSrc2dDescriptor(Lengths_, gammaStrides_, numBlockTileIteration_);
beta_grid_desc_m_k_ = beta_grid_desc_m_k_ =
MakeSrc2dDescriptor(Lengths_, betaStrides_, blkGroupSize_, numBlockTileIteration_); MakeSrc2dDescriptor(Lengths_, betaStrides_, numBlockTileIteration_);
y_grid_desc_m_k_ = y_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, yStrides_, numBlockTileIteration_);
MakeSrc2dDescriptor(Lengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
isSweeponce_ = isSweeponce_ =
x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize; x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
...@@ -202,7 +199,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType, ...@@ -202,7 +199,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
YElementwiseOperation y_elementwise_op_; YElementwiseOperation y_elementwise_op_;
int blkGroupSize_;
int numBlockTileIteration_; int numBlockTileIteration_;
size_t gridSize_; size_t gridSize_;
...@@ -286,6 +282,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType, ...@@ -286,6 +282,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0) if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
return false; return false;
if(p_arg_->invariant_lowest_length % YDstVectorSize != 0)
return false;
}; };
} }
else else
...@@ -295,12 +294,12 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType, ...@@ -295,12 +294,12 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0) if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
return false; return false;
};
if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0) if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
{ {
return false; return false;
} }
};
// if fastest dim is not reduced // if fastest dim is not reduced
if constexpr(GammaSrcVectorDim == 0) if constexpr(GammaSrcVectorDim == 0)
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 7546eb7a
...@@ -56,6 +56,12 @@ struct PassThrough ...@@ -56,6 +56,12 @@ struct PassThrough
y = type_convert<bhalf_t>(x); y = type_convert<bhalf_t>(x);
} }
template <>
__host__ __device__ void operator()<bhalf_t, half_t>(bhalf_t& y, const half_t& x) const
{
y = type_convert<bhalf_t>(x);
}
template <> template <>
__host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const __host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
{ {
...@@ -86,6 +92,23 @@ struct UnaryConvert ...@@ -86,6 +92,23 @@ struct UnaryConvert
} }
}; };
struct ConvertBF16RTN
{
// convert to bf16 using round to nearest (rtn)
template <typename Y, typename X>
__host__ __device__ void operator()(Y& y, const X& x) const
{
// check Y datatype
static_assert(is_same<Y, bhalf_t>::value, "Data type is not supported by this operation!");
// check X datatype
static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
"Data type is not supported by this operation!");
y = bf16_convert_rtn<Y>(x);
}
};
struct Scale struct Scale
{ {
__host__ __device__ Scale(float scale) : scale_(scale) {} __host__ __device__ Scale(float scale) : scale_(scale) {}
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
View file @ 7546eb7a
...@@ -96,7 +96,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle ...@@ -96,7 +96,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update // when mfma if fixed, remove this section and update
// ABDataTypeAdjusted -> ABDataType throughout this file // ABDataTypeAdjusted -> ABDataType throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__) #if CK_WORKAROUND_DENORM_FIX
using ABDataTypeAdjusted = using ABDataTypeAdjusted =
conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>; conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
#else #else
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
View file @ 7546eb7a
...@@ -266,7 +266,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight ...@@ -266,7 +266,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update // when mfma if fixed, remove this section and update
// FloatABAdjusted -> FloatAB throughout this file // FloatABAdjusted -> FloatAB throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__) #if CK_WORKAROUND_DENORM_FIX
using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>; using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
#else #else
using FloatABAdjusted = FloatAB; using FloatABAdjusted = FloatAB;
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ 7546eb7a
...@@ -136,7 +136,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 ...@@ -136,7 +136,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update // when mfma if fixed, remove this section and update
// FloatABAdjusted -> FloatAB throughout this file // FloatABAdjusted -> FloatAB throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__) #if CK_WORKAROUND_DENORM_FIX
using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>; using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
#else #else
using FloatABAdjusted = FloatAB; using FloatABAdjusted = FloatAB;
......
include/ck/tensor_operation/gpu/grid/gridwise_normalization_selector.hpp → include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
View file @ 7546eb7a
...@@ -3,8 +3,8 @@ ...@@ -3,8 +3,8 @@
#pragma once #pragma once
#include "ck/tensor_operation/gpu/grid/gridwise_normalization_naive_variance.hpp" #include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_normalization_welford_variance.hpp" #include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp"
namespace ck { namespace ck {
template <typename GridwiseReduction, template <typename GridwiseReduction,
......
include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp 0 → 100644
View file @ 7546eb7a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/math.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
template <typename XDataType,
typename ComputeDataType,
typename MeanVarDataType,
typename XGridDesc_M_K,
typename MeanVarGridDesc_M_KBlock,
index_t BlockSize,
index_t MThreadClusterSize,
index_t KThreadClusterSize,
index_t MThreadSliceSize,
index_t KThreadSliceSize,
index_t XSrcVectorDim,
index_t XSrcVectorSize>
struct GridwiseNormalizationSplitK1st
{
static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
(XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
"Invalid thread slice sizes and/or vector sizes configuration, please check!");
static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
using ThreadBufferDimAccessOrder =
typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
using ThreadClusterArrangeOrder =
typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
static constexpr auto thread_cluster_desc =
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
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));
using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
using ThreadReduceDstDesc_M =
decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
using ThreadwiseWelford =
ThreadwiseWelford<ComputeDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
BlockSize,
ThreadClusterLengths_M_K,
ThreadClusterArrangeOrder,
false>;
using PassThroughOp = tensor_operation::element_wise::PassThrough;
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
__device__ static int
GetKPerThread(int kRaw, int kGridSize, int block_k_cluster_id, int thread_k_cluster_id)
{
bool is_rightmost_block = block_k_cluster_id == kGridSize - 1;
if(is_rightmost_block)
{
int left_kPerBlock = math::integer_divide_ceil(kRaw, kGridSize);
int kPerBlock = kRaw % kGridSize == 0 ? left_kPerBlock : kRaw % left_kPerBlock;
int kPerThread =
kPerBlock < K_BlockTileSize ? 0 : KThreadSliceSize * (kPerBlock / K_BlockTileSize);
int kPerBlockTail = kPerBlock - kPerThread * KThreadClusterSize;
if(kPerBlockTail > 0)
{
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
int thread_max_len =
(thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
int delta = thread_max_len - kPerBlockTail;
delta = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
kPerThread += XSrcVectorSize - delta;
});
}
return kPerThread;
}
else
{
int kPerBlock = math::integer_divide_ceil(kRaw, kGridSize);
return KThreadSliceSize * (kPerBlock / K_BlockTileSize);
}
}
// Calculate mean and variance by welford along k dimension
__device__ static void Run(const XGridDesc_M_K& x_grid_desc_m_k,
const MeanVarGridDesc_M_KBlock& mean_var_grid_desc_m_kblock,
index_t num_k_block_tile_iteration,
const XDataType* const __restrict__ p_x_global,
MeanVarDataType* const p_mean_global,
MeanVarDataType* const p_variance_global,
int32_t* const p_welford_count_global)
{
auto x_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
ComputeDataType,
MThreadSliceSize * XSrcVectorSize,
true>{};
},
Number<ThreadBufferNumber>{});
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
var_thread_buf;
const index_t thread_local_id = get_thread_local_1d_id();
const index_t block_global_id = get_block_1d_id();
const index_t k_grid_size = mean_var_grid_desc_m_kblock.GetLength(I1);
const index_t block_m_cluster_id = block_global_id / k_grid_size;
const index_t block_k_cluster_id = block_global_id % k_grid_size;
const auto thread_cluster_idx =
thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_k_cluster_id = thread_cluster_idx[I1];
const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
ComputeDataType,
XGridDesc_M_K,
decltype(thread_buffer_desc_m_k),
ThreadBufferLengths_M_K,
ThreadBufferDimAccessOrder,
XSrcVectorDim,
XSrcVectorSize,
1,
true>(
x_grid_desc_m_k,
make_multi_index(
block_m_cluster_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
block_k_cluster_id * reduceSizePerBlock + thread_k_cluster_id * XSrcVectorSize));
auto mean_var_count_store_index = make_multi_index(
block_m_cluster_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
block_k_cluster_id);
auto threadwise_welford_mean_var_store =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
MeanVarDataType,
decltype(thread_buffer_desc_m_1),
MeanVarGridDesc_M_KBlock,
PassThroughOp,
ThreadBufferLengths_M_1,
Sequence<0, 1>,
1,
1,
InMemoryDataOperationEnum::Set,
1,
true>(
mean_var_grid_desc_m_kblock, mean_var_count_store_index, PassThroughOp{});
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
auto mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_mean_global, mean_var_grid_desc_m_kblock.GetElementSpaceSize());
auto var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_variance_global, mean_var_grid_desc_m_kblock.GetElementSpaceSize());
auto threadwise_welford = ThreadwiseWelford();
int kRaw = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
threadwise_welford.max_count_ =
GetKPerThread(kRaw, k_grid_size, block_k_cluster_id, thread_k_cluster_id);
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
mean_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
var_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
});
for(index_t k = 0; k < num_k_block_tile_iteration; ++k)
{
static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
});
}
int welford_count = 0;
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
if constexpr(I > 0)
block_sync_lds();
int count = threadwise_welford.cur_count_;
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
// The value of count is same for all I
if constexpr(I == MThreadSliceSize - 1)
welford_count = count;
});
if(thread_k_cluster_id == 0)
{
threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
mean_thread_buf,
mean_var_grid_desc_m_kblock,
mean_global_val_buf);
threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
make_tuple(I0, I0),
var_thread_buf,
mean_var_grid_desc_m_kblock,
var_global_val_buf);
if(block_m_cluster_id == 0 && thread_m_cluster_id == 0)
p_welford_count_global[block_k_cluster_id] = welford_count;
}
}
};
} // namespace ck
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