ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-06 15:54:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

Introduce gemm_softmax_gemm to codegen.

Browse Source
This commit is contained in:
Mirza Halilcevic
2024-09-25 08:22:07 +00:00
parent 3528a523ff
commit d43cd4ad32
52 changed files with 2108 additions and 187 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
include/ck/ck.hpp
View File

@@ -4,16 +4,12 @@
#pragma once
#include "ck/config.h"
#include "ck/utility/env.hpp"
#ifndef __HIPCC_RTC__
#ifndef CK_DONT_USE_HIP_RUNTIME_HEADERS
#include "hip/hip_runtime.h"
#include "hip/hip_fp16.h"
#endif
// environment variable to enable logging:
// export CK_LOGGING=ON or CK_LOGGING=1 or CK_LOGGING=ENABLED
CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
#endif
// to do: add various levels of logging with CK_LOG_LEVEL

2
include/ck/host_utility/device_prop.hpp
View File

@@ -3,6 +3,7 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <string>
#include <map>
#include <hip/hip_runtime.h>
@@ -96,3 +97,4 @@ inline bool is_gfx12_supported()
}
} // namespace ck
#endif

3
include/ck/host_utility/kernel_launch.hpp
View File

@@ -2,7 +2,7 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef __HIPCC_RTC__
#include <hip/hip_runtime.h>
#include "ck/ck.hpp"
@@ -160,3 +160,4 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
return 0;
#endif
}
#endif

8
include/ck/tensor_operation/gpu/device/device_base.hpp
View File

@@ -3,15 +3,17 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <string>
#include <sstream>
#include "ck/stream_config.hpp"
#endif
namespace ck {
namespace tensor_operation {
namespace device {
#ifndef __HIPCC_RTC__
struct BaseArgument
{
BaseArgument() = default;
@@ -36,6 +38,7 @@ struct BaseInvoker
virtual ~BaseInvoker() {}
};
#endif
struct BaseOperator
{
@@ -43,6 +46,7 @@ struct BaseOperator
BaseOperator(const BaseOperator&) = default;
BaseOperator& operator=(const BaseOperator&) = default;
#ifndef __HIPCC_RTC__
virtual bool IsSupportedArgument(const BaseArgument*) { return false; }
virtual std::string GetTypeString() const { return ""; }
@@ -66,7 +70,7 @@ struct BaseOperator
assert(p_arg);
p_arg->p_workspace_ = p_workspace;
}
#endif
virtual ~BaseOperator() {}
};

5
include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
View File

@@ -2,9 +2,10 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef __HIPCC_RTC__
#include <iostream>
#include <vector>
#endif
#include "device_base.hpp"
@@ -28,6 +29,7 @@ template <typename ALayout,
bool MaskOutUpperTriangle> // TODO: enum for mask type
struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
{
#ifndef __HIPCC_RTC__
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b0,
@@ -53,6 +55,7 @@ struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
CElementwiseOperation c_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
#endif
};
} // namespace device

8
include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
View File

@@ -2,9 +2,11 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#ifndef __HIPCC_RTC__
#include <array>
#endif
#include "ck/utility/array.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
@@ -34,6 +36,7 @@ struct DeviceGemmMultipleD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
#ifndef __HIPCC_RTC__
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
@@ -51,6 +54,7 @@ struct DeviceGemmMultipleD : public BaseOperator
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
#endif
};
// GEMM:
@@ -76,6 +80,7 @@ struct DeviceGemmMultipleDSplitK : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
#ifndef __HIPCC_RTC__
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
@@ -94,6 +99,7 @@ struct DeviceGemmMultipleDSplitK : public BaseOperator
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
#endif
};
} // namespace device

3
include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
View File

@@ -28,7 +28,7 @@ enum struct GemmSpecialization
NKOPadding,
MNKOPadding,
};
#ifndef __HIPCC_RTC__
inline std::string getGemmSpecializationString(const GemmSpecialization& s)
{
switch(s)
@@ -52,6 +52,7 @@ inline std::string getGemmSpecializationString(const GemmSpecialization& s)
default: return "Unrecognized specialization!";
}
}
#endif
} // namespace device
} // namespace tensor_operation

404
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
View File

@@ -3,8 +3,12 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <iostream>
#include <sstream>
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#endif
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
@@ -15,8 +19,6 @@
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
@@ -40,27 +42,27 @@ template <typename GridwiseGemm,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op,
const B1ElementwiseOperation b1_element_op,
const CElementwiseOperation c_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap block_2_ctile_map,
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask)
kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op,
const B1ElementwiseOperation b1_element_op,
const CElementwiseOperation c_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap block_2_ctile_map,
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
const C0MatrixMask c0_matrix_mask)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
@@ -430,6 +432,7 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
matrix_padder.PadN,
MaskOutUpperTriangle>;
#ifndef __HIPCC_RTC__
// Argument
struct Argument : public BaseArgument
{
@@ -604,6 +607,7 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
#endif
static constexpr bool IsValidCompilationParameter()
{
@@ -611,6 +615,97 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return true;
}
static constexpr bool
IsSupported(index_t MRaw_, index_t NRaw_, index_t KRaw_, index_t Gemm1NRaw_)
{
// check vector load/store
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
// check vector load of A
if constexpr(is_same_v<ALayout, Row>)
{
if(KRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<ALayout, Col>)
{
if(MRaw_ % ABlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector load of B
if constexpr(is_same_v<BLayout, Row>)
{
if(NRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<BLayout, Col>)
{
if(KRaw_ % BBlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector load of B1
if constexpr(is_same_v<B1Layout, Row>)
{
if(Gemm1NRaw_ % B1BlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else if constexpr(is_same_v<B1Layout, Col>)
{
if(NRaw_ % B1BlockTransferSrcScalarPerVector != 0)
{
return false;
}
}
else
{
return false;
}
// check vector load of C
if constexpr(is_same_v<CLayout, Row>)
{
if(Gemm1NRaw_ % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
return false;
}
}
else if constexpr(is_same_v<CLayout, Col>)
{
if(MRaw_ % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
{
return false;
}
}
else
{
return false;
}
return true;
}
#ifndef __HIPCC_RTC__
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
@@ -765,8 +860,271 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return str.str();
}
#endif
template <class ADesc, class BDesc, class B1Desc, class CDesc>
struct Descriptor
{
template <class AGridDescriptor>
static constexpr auto MakeAGridDescriptor_AK0_M_AK1(const AGridDescriptor& a_grid_desc)
{
const auto a_grid_desc_m_k = DeviceOp::matrix_padder.PadADescriptor_M_K(a_grid_desc);
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
const auto AK0 = K / AK1;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
template <class BGridDescriptor>
static constexpr auto MakeBGridDescriptor_BK0_N_BK1(const BGridDescriptor& b_grid_desc)
{
const auto b_grid_desc_n_k = DeviceOp::matrix_padder.PadBDescriptor_N_K(b_grid_desc);
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
const auto BK0 = K / BK1;
return transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
template <class B1GridDescriptor>
static constexpr auto MakeB1GridDescriptor_BK0_N_BK1(const B1GridDescriptor& b1_grid_desc)
{
const auto b1_grid_desc_n_k = DeviceOp::matrix_padder.PadB1Descriptor_N_K(b1_grid_desc);
const auto N = b1_grid_desc_n_k.GetLength(I0);
const auto K = b1_grid_desc_n_k.GetLength(I1);
const auto B1K0 = K / B1K1;
return transform_tensor_descriptor(
b1_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
template <class CGridDescriptor>
static constexpr auto MakeCGridDescriptor_M_N(const CGridDescriptor& c_grid_desc)
{
return DeviceOp::matrix_padder.PadCDescriptor_M_N(c_grid_desc);
}
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(MakeAGridDescriptor_AK0_M_AK1(ADesc{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1(BDesc{}))>;
using B1GridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(MakeB1GridDescriptor_BK0_N_BK1(B1Desc{}))>;
using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(CDesc{}))>;
// GridwiseGemm
using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
B1GridDesc_BK0_N_BK1,
CGridDesc_M_N,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
Gemm1NPerBlock,
Gemm1KPerBlock,
AK1,
BK1,
B1K1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
Gemm1NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
true,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
true,
BBlockLdsExtraN,
B1BlockTransferThreadClusterLengths_BK0_N_BK1,
B1BlockTransferThreadClusterArrangeOrder,
B1BlockTransferSrcAccessOrder,
B1BlockTransferSrcVectorDim,
B1BlockTransferSrcScalarPerVector,
B1BlockTransferDstScalarPerVector_BK1,
false,
B1BlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopSched,
matrix_padder.PadN,
MaskOutUpperTriangle>;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1;
CGridDesc_M_N c_grid_desc_m_n;
C0MatrixMask c0_matrix_mask;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_descriptor_mblock_mperblock_nblock_nperblock;
// element-wise op
AElementwiseOperation a_element_op;
BElementwiseOperation b_element_op;
B1ElementwiseOperation b1_element_op;
CElementwiseOperation c_element_op;
bool has_main_k_block_loop = true;
bool is_valid = false;
constexpr Descriptor(ADesc a,
BDesc b,
B1Desc b1,
CDesc c,
AElementwiseOperation a_element_op_,
BElementwiseOperation b_element_op_,
B1ElementwiseOperation b1_element_op_,
CElementwiseOperation c_element_op_)
: a_grid_desc_ak0_m_ak1{MakeAGridDescriptor_AK0_M_AK1(a)},
b_grid_desc_bk0_n_bk1{MakeBGridDescriptor_BK0_N_BK1(b)},
b1_grid_desc_bk0_n_bk1{MakeB1GridDescriptor_BK0_N_BK1(b1)},
c_grid_desc_m_n{MakeCGridDescriptor_M_N(c)},
block_2_ctile_map{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n)},
c_grid_descriptor_mblock_mperblock_nblock_nperblock{
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n)},
has_main_k_block_loop{GridwiseGemm::CalculateHasMainKBlockLoop(
a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2))},
c0_matrix_mask{c.GetLength(I1)},
a_element_op{a_element_op_},
b_element_op{b_element_op_},
b1_element_op{b1_element_op_},
c_element_op{c_element_op_},
is_valid{GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
b1_grid_desc_bk0_n_bk1,
c_grid_desc_m_n,
block_2_ctile_map) and
IsSupported(a_grid_desc_ak0_m_ak1.GetLength(I1),
b_grid_desc_bk0_n_bk1.GetLength(I1),
a_grid_desc_ak0_m_ak1.GetLength(I0) *
a_grid_desc_ak0_m_ak1.GetLength(I2),
b1_grid_desc_bk0_n_bk1.GetLength(I1))}
{
}
constexpr bool IsValid() const { return is_valid; }
};
template <class ADesc, class BDesc, class B1Desc, class CDesc>
static constexpr auto
make_descriptor(ADesc a,
BDesc b,
B1Desc b1,
CDesc c,
AElementwiseOperation a_element_op = AElementwiseOperation{},
BElementwiseOperation b_element_op = BElementwiseOperation{},
B1ElementwiseOperation b1_element_op = B1ElementwiseOperation{},
CElementwiseOperation c_element_op = CElementwiseOperation{})
{
return Descriptor<ADesc, BDesc, B1Desc, CDesc>(
a, b, b1, c, a_element_op, b_element_op, b1_element_op, c_element_op);
}
template <class Desc>
__device__ static void Run(const Desc& desc,
const float scale,
const ADataType* __restrict__ p_a_grid,
const ADataType* __restrict__ p_b_grid,
const ADataType* __restrict__ p_b1_grid,
CDataType* __restrict__ p_c_grid)
{
#ifndef __HIPCC_RTC__
assert(desc.is_valid);
#endif
__shared__ char p_shared_block[Desc::GridwiseGemm::GetSharedMemoryNumberOfByte()];
AccElementwiseOperation acc_element_op{scale};
if(desc.has_main_k_block_loop)
{
Desc::GridwiseGemm::template Run<true>(
p_a_grid,
p_b_grid,
p_b1_grid,
p_c_grid,
p_shared_block,
desc.a_element_op,
desc.b_element_op,
acc_element_op,
desc.b1_element_op,
desc.c_element_op,
desc.a_grid_desc_ak0_m_ak1,
desc.b_grid_desc_bk0_n_bk1,
desc.b1_grid_desc_bk0_n_bk1,
desc.c_grid_descriptor_mblock_mperblock_nblock_nperblock,
desc.block_2_ctile_map,
desc.c0_matrix_mask);
}
else
{
Desc::GridwiseGemm::template Run<false>(
p_a_grid,
p_b_grid,
p_b1_grid,
p_c_grid,
p_shared_block,
desc.a_element_op,
desc.b_element_op,
acc_element_op,
desc.b1_element_op,
desc.c_element_op,
desc.a_grid_desc_ak0_m_ak1,
desc.b_grid_desc_bk0_n_bk1,
desc.b1_grid_desc_bk0_n_bk1,
desc.c_grid_descriptor_mblock_mperblock_nblock_nperblock,
desc.block_2_ctile_map,
desc.c0_matrix_mask);
}
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
} // namespace ck

57
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
View File

@@ -3,8 +3,12 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <iostream>
#include <sstream>
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#endif
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
@@ -14,8 +18,6 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
@@ -35,22 +37,22 @@ template <typename GridwiseGemm,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_multiple_d_xdl_cshuffle(const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
kernel_gemm_multiple_d_xdl_cshuffle(const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
@@ -225,9 +227,9 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
const std::array<index_t, NumDTensor>& NRaws,
const std::array<index_t, NumDTensor>& DsStride)
static auto MakeDsGridDescriptor_M_N(const Array<index_t, NumDTensor>& MRaws,
const Array<index_t, NumDTensor>& NRaws,
const Array<index_t, NumDTensor>& DsStride)
{
return generate_tuple(
[&](auto i) {
@@ -309,6 +311,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
using Block2ETileMap =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
#ifndef __HIPCC_RTC__
// Argument
struct Argument : public BaseArgument
{
@@ -498,6 +501,8 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
}
};
#endif
static constexpr bool IsSupported(index_t MRaw_, index_t NRaw_, index_t KRaw_)
{
// check vector load/store
@@ -578,6 +583,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
return true;
}
#ifndef __HIPCC_RTC__
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
@@ -676,11 +682,13 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
{
auto str = std::stringstream();
std::map<LoopScheduler, std::string> LoopSchedToString{
{LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
std::map<LoopScheduler, std::string> LoopSchedToString{{LoopScheduler::Default, "Default"},
{ LoopScheduler::Interwave,
"Interwave" }};
std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
{PipelineVersion::v2, "v2"}};
{ PipelineVersion::v2,
"v2" }};
// clang-format off
str << "DeviceGemmMultipleD_Xdl_CShuffle"
@@ -709,6 +717,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
return str.str();
}
#endif
template <class ADesc, class BDesc, class DsDesc, class EDesc>
struct Descriptor
@@ -847,7 +856,9 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
EDataType* __restrict__ p_e_grid)
{
__shared__ char p_shared_block[GridwiseGemm::GetSharedMemoryNumberOfByte()];
#ifndef __HIPCC_RTC__
assert(desc.IsValid());
#endif
if(desc.has_main_k_block_loop)
{
GridwiseGemm::template Run<true>(p_a_grid,

4
include/ck/tensor_operation/gpu/device/masking_specialization.hpp
View File

@@ -13,6 +13,7 @@ enum struct MaskingSpecialization
MaskOutUpperTriangle
};
#ifndef __HIPCC_RTC__
inline std::string getMaskingSpecializationString(const MaskingSpecialization& s)
{
switch(s)
@@ -22,6 +23,7 @@ inline std::string getMaskingSpecializationString(const MaskingSpecialization& s
default: return "Unrecognized specialization!";
}
}
#endif
struct MaskDisabledPredicate
{
@@ -53,7 +55,7 @@ struct MaskOutUpperTrianglePredicate
template <typename MaskOutPredicate>
struct C0MatrixMask_impl
{
__host__ __device__ C0MatrixMask_impl(index_t NRaw)
__host__ __device__ constexpr C0MatrixMask_impl(index_t NRaw)
: NRaw_(NRaw), predicate_(MaskOutPredicate{})
{
}

2
include/ck/tensor_operation/gpu/device/tensor_layout.hpp
View File

@@ -430,6 +430,7 @@ struct G_NDHW : public BaseTensorLayout
} // namespace convolution
#ifndef __HIPCC_RTC__
template <
typename Layout,
typename std::enable_if<std::is_base_of<BaseTensorLayout, Layout>::value, bool>::type = false>
@@ -438,6 +439,7 @@ std::ostream& operator<<(std::ostream& os, const Layout&)
os << Layout::name;
return os;
}
#endif
} // namespace tensor_layout
} // namespace ck

4
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View File

@@ -340,8 +340,8 @@ struct Bilinear
};
template <>
__host__ __device__ constexpr void operator()<std::int8_t, std::int32_t, std::int8_t>(
std::int8_t& y, const std::int32_t& x0, const std::int8_t& x1) const
__host__ __device__ constexpr void operator()<int8_t, int32_t, int8_t>(
int8_t& y, const int32_t& x0, const int8_t& x1) const
{
y = type_convert<int8_t>(alpha_ * type_convert<float>(x0) +
beta_ * type_convert<float>(x1));

4
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View File

@@ -466,7 +466,7 @@ struct FastGelu
template <typename Y, typename X>
__device__ void operator()(Y& y, const X& x) const;
#ifndef __HIPCC_RTC__
template <>
__host__ void operator()<float, float>(float& y, const float& x) const
{
@@ -477,7 +477,7 @@ struct FastGelu
const float emu = exp(u);
y = x / (1.f + emu);
}
#endif
// device code, use lower precision "__ocml_exp_f32" and "rcp"
template <>
__device__ void operator()<float, float>(float& y, const float& x) const

6
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View File

@@ -7,8 +7,10 @@
#include "ck/utility/number.hpp"
#include "ck/tensor_description/tensor_adaptor.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#ifndef __HIPCC_RTC__
#include <limits>
#include <stdlib.h>
#endif
namespace ck {
@@ -979,7 +981,7 @@ struct BlockToCTileMap_3DGrid_KSplit
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);
return ck::make_tuple(N0, M0, k_split);
}
template <typename TopIdx>
@@ -1103,7 +1105,7 @@ struct BlockToCTileMap_GemmStreamK
uint32_t dp_for_sk_iters = k_iters_per_tile.get();
uint32_t best_sk_score =
std::numeric_limits<int>::max(); // we need to find the smallest sk iters
ck::NumericLimits<int>::Max(); // we need to find the smallest sk iters
for(uint32_t tentative_sk_blocks = min_sk_tiles; tentative_sk_blocks < max_sk_tiles;
tentative_sk_blocks++)
{

8
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
View File

@@ -475,9 +475,9 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
template <typename DsLayout, GemmSpecialization GemmSpec>
__host__ __device__ static auto
MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
const std::array<index_t, NumDTensor>& NRaws,
const std::array<index_t, NumDTensor>& DsStride)
MakeDsGridDescriptor_M_N(const Array<index_t, NumDTensor>& MRaws,
const Array<index_t, NumDTensor>& NRaws,
const Array<index_t, NumDTensor>& DsStride)
{
return generate_tuple(
[&](auto i) {
@@ -941,7 +941,7 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
const index_t K,
const index_t StrideA,
const index_t StrideB,
const std::array<index_t, NumDTensor> StrideDs,
const Array<index_t, NumDTensor> StrideDs,
const index_t StrideE,
const Block2ETileMap& block_2_etile_map)
{

6
include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
View File

@@ -3,8 +3,10 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <iostream>
#include <ostream>
#endif
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
@@ -53,12 +55,15 @@ constexpr auto GridwiseGemmPipeline_Selector()
}
else
{
#ifndef __HIPCC_RTC__
std::cerr << "GridwiseGemmPipeline configuration is not available" << std::endl;
#endif
}
}
} // namespace ck
#ifndef __HIPCC_RTC__
inline std::ostream& operator<<(std::ostream& os, const ck::PipelineVersion& p)
{
switch(p)
@@ -71,3 +76,4 @@ inline std::ostream& operator<<(std::ostream& os, const ck::PipelineVersion& p)
}
return os;
}
#endif

2
include/ck/utility/amd_buffer_addressing.hpp
View File

@@ -1005,6 +1005,7 @@ llvm_amdgcn_raw_buffer_load_lds(int32x4_t rsrc,
index_t offset,
index_t aux) __asm("llvm.amdgcn.raw.buffer.load.lds");
#ifndef __HIPCC_RTC__
template <typename T, index_t NumElemsPerThread>
__device__ void amd_direct_load_global_to_lds(const T* global_base_ptr,
const index_t global_offset,
@@ -1042,5 +1043,6 @@ __device__ void amd_direct_load_global_to_lds(const T* global_base_ptr,
src_resource, lds_ptr, sizeof(uint32_t), global_offset_bytes, 0, 0, 0);
#endif
}
#endif
} // namespace ck

22
include/ck/utility/amd_wave_read_first_lane.hpp
View File

@@ -7,10 +7,12 @@
#include "ck/utility/functional2.hpp"
#include "ck/utility/math.hpp"
#ifndef __HIPCC_RTC__
#include <array>
#include <cstddef>
#include <cstdint>
#include <type_traits>
#endif
namespace ck {
namespace detail {
@@ -37,7 +39,7 @@ struct get_carrier<3>
{
using value_type = uint32_t;
std::array<std::byte, 3> bytes;
Array<ck::byte, 3> bytes;
static_assert(sizeof(bytes) <= sizeof(value_type));
// replacement of host std::copy_n()
@@ -61,22 +63,22 @@ struct get_carrier<3>
// method to trigger template substitution failure
__device__ carrier(const carrier& other) noexcept
{
copy_n(other.bytes.begin(), bytes.size(), bytes.begin());
copy_n(other.bytes.begin(), bytes.Size(), bytes.begin());
}
public:
__device__ carrier& operator=(value_type value) noexcept
{
copy_n(reinterpret_cast<const std::byte*>(&value), bytes.size(), bytes.begin());
copy_n(reinterpret_cast<const ck::byte*>(&value), bytes.Size(), bytes.begin());
return *this;
}
__device__ operator value_type() const noexcept
{
std::byte result[sizeof(value_type)];
ck::byte result[sizeof(value_type)];
copy_n(bytes.begin(), bytes.size(), result);
copy_n(bytes.begin(), bytes.Size(), result);
return *reinterpret_cast<const value_type*>(result);
}
@@ -109,8 +111,8 @@ __device__ inline int64_t amd_wave_read_first_lane(int64_t value)
{
constexpr unsigned object_size = sizeof(int64_t);
constexpr unsigned second_part_offset = object_size / 2;
auto* const from_obj = reinterpret_cast<const std::byte*>(&value);
alignas(int64_t) std::byte to_obj[object_size];
auto* const from_obj = reinterpret_cast<const ck::byte*>(&value);
alignas(int64_t) ck::byte to_obj[object_size];
using Sgpr = uint32_t;
@@ -124,15 +126,15 @@ __device__ inline int64_t amd_wave_read_first_lane(int64_t value)
template <
typename Object,
typename = std::enable_if_t<std::is_class_v<Object> && std::is_trivially_copyable_v<Object>>>
typename = ck::enable_if_t<ck::is_class_v<Object> && ck::is_trivially_copyable_v<Object>>>
__device__ auto amd_wave_read_first_lane(const Object& obj)
{
using Size = unsigned;
constexpr Size SgprSize = 4;
constexpr Size ObjectSize = sizeof(Object);
auto* const from_obj = reinterpret_cast<const std::byte*>(&obj);
alignas(Object) std::byte to_obj[ObjectSize];
auto* const from_obj = reinterpret_cast<const ck::byte*>(&obj);
alignas(Object) ck::byte to_obj[ObjectSize];
constexpr Size RemainedSize = ObjectSize % SgprSize;
constexpr Size CompleteSgprCopyBoundary = ObjectSize - RemainedSize;

4
include/ck/utility/array.hpp
View File

@@ -38,6 +38,8 @@ struct Array
}
__host__ __device__ constexpr const TData* begin() const { return &mData[0]; }
__host__ __device__ constexpr const TData* end() const { return &mData[NSize]; }
__host__ __device__ constexpr TData* begin() { return &mData[0]; }
__host__ __device__ constexpr TData* end() { return &mData[NSize]; }
};
// empty Array
@@ -54,7 +56,7 @@ template <typename X, typename... Xs>
__host__ __device__ constexpr auto make_array(X&& x, Xs&&... xs)
{
using data_type = remove_cvref_t<X>;
return Array<data_type, sizeof...(Xs) + 1>{std::forward<X>(x), std::forward<Xs>(xs)...};
return Array<data_type, sizeof...(Xs) + 1>{ck::forward<X>(x), ck::forward<Xs>(xs)...};
}
// make empty array

4
include/ck/utility/container_helper.hpp
View File

@@ -326,14 +326,14 @@ template <typename T, index_t NX, index_t NY>
__host__ __device__ constexpr auto container_concat(const Array<T, NX>& ax, const Array<T, NY>& ay)
{
return unpack2(
[&](auto&&... zs) { return make_array(std::forward<decltype(zs)>(zs)...); }, ax, ay);
[&](auto&&... zs) { return make_array(ck::forward<decltype(zs)>(zs)...); }, ax, ay);
}
template <typename... X, typename... Y>
__host__ __device__ constexpr auto container_concat(const Tuple<X...>& tx, const Tuple<Y...>& ty)
{
return unpack2(
[&](auto&&... zs) { return make_tuple(std::forward<decltype(zs)>(zs)...); }, tx, ty);
[&](auto&&... zs) { return make_tuple(ck::forward<decltype(zs)>(zs)...); }, tx, ty);
}
template <typename Container>

158
include/ck/utility/data_type.hpp
View File

@@ -5,8 +5,25 @@
#include "ck/utility/statically_indexed_array.hpp"
#ifdef __HIPCC_RTC__
/// Definitions from <cstdint>, <cmath> conflict with
/// /opt/rocm/include/hip/amd_detail/amd_hip_vector_types.h.
using int8_t = signed char;
using uint8_t = unsigned char;
using int16_t = signed short;
using uint16_t = unsigned short;
using float_t = float;
#endif // __HIPCC_RTC__
namespace ck {
#ifdef __HIPCC_RTC__
using byte = unsigned char;
#else
using std::byte;
#endif
using bhalf_t = ushort;
using half_t = _Float16;
using int4_t = _BitInt(4);
@@ -1060,6 +1077,146 @@ using uint8x16_t = typename vector_type<uint8_t, 16>::type;
using uint8x32_t = typename vector_type<uint8_t, 32>::type;
using uint8x64_t = typename vector_type<uint8_t, 64>::type;
#ifdef __HIPCC_RTC__
template <typename T>
struct NumericLimits;
template <>
struct NumericLimits<int32_t>
{
__host__ __device__ static constexpr int32_t Lowest() noexcept { return -2147483647 - 1; }
__host__ __device__ static constexpr int32_t Min() noexcept { return -2147483647 - 1; }
__host__ __device__ static constexpr int32_t Max() noexcept { return 2147483647; }
__host__ __device__ static constexpr int32_t Infinity() noexcept { return 0; }
__host__ __device__ static constexpr int32_t QuietNaN() { return 0; }
};
template <>
struct NumericLimits<int16_t>
{
__host__ __device__ static constexpr int16_t Lowest() noexcept { return -32768; }
__host__ __device__ static constexpr int16_t Min() noexcept { return -32768; }
__host__ __device__ static constexpr int16_t Max() noexcept { return 32767; }
__host__ __device__ static constexpr int16_t Infinity() noexcept { return 0; }
__host__ __device__ static constexpr int16_t QuietNaN() { return 0; }
};
template <>
struct NumericLimits<int8_t>
{
__host__ __device__ static constexpr int8_t Lowest() noexcept { return -128; }
__host__ __device__ static constexpr int8_t Min() noexcept { return -128; }
__host__ __device__ static constexpr int8_t Max() noexcept { return 127; }
__host__ __device__ static constexpr int8_t Infinity() noexcept { return 0; }
__host__ __device__ static constexpr int8_t QuietNaN() { return 0; }
};
template <>
struct NumericLimits<uint32_t>
{
__host__ __device__ static constexpr uint32_t Lowest() noexcept { return 0; }
__host__ __device__ static constexpr uint32_t Min() noexcept { return 0; }
__host__ __device__ static constexpr uint32_t Max() noexcept { return 4294967295U; }
__host__ __device__ static constexpr uint32_t Infinity() noexcept { return 0; }
__host__ __device__ static constexpr uint32_t QuietNaN() { return 0; }
};
template <>
struct NumericLimits<uint16_t>
{
__host__ __device__ static constexpr uint16_t Lowest() noexcept { return 0; }
__host__ __device__ static constexpr uint16_t Min() noexcept { return 0; }
__host__ __device__ static constexpr uint16_t Max() noexcept { return 65535U; }
__host__ __device__ static constexpr uint16_t Infinity() noexcept { return 0; }
__host__ __device__ static constexpr uint16_t QuietNaN() { return 0; }
};
template <>
struct NumericLimits<float>
{
static constexpr unsigned int binary_min = 0x00800000;
static constexpr unsigned int binary_max = 0x7F7FFFFF;
static constexpr unsigned int binary_lowest = 0xFF7FFFFF;
static constexpr unsigned int binary_qnan = 0xFFC00001;
static constexpr unsigned int binary_inf = 0x7F8000000;
__host__ __device__ static constexpr float Min() { return bit_cast<float>(binary_min); }
__host__ __device__ static constexpr float Max() { return bit_cast<float>(binary_max); }
__host__ __device__ static constexpr float Lowest() { return bit_cast<float>(binary_lowest); }
__host__ __device__ static constexpr float QuietNaN() { return bit_cast<float>(binary_qnan); }
__host__ __device__ static constexpr float Infinity() { return bit_cast<float>(binary_inf); }
};
template <>
struct NumericLimits<half_t>
{
static constexpr unsigned short binary_min = 0x0400;
static constexpr unsigned short binary_max = 0x7BFF;
static constexpr unsigned short binary_lowest = 0xFBFF;
static constexpr unsigned short binary_qnan = 0x7FFF;
__host__ __device__ static constexpr half_t Min() { return bit_cast<half_t>(binary_min); }
__host__ __device__ static constexpr half_t Max() { return bit_cast<half_t>(binary_max); }
__host__ __device__ static constexpr half_t Lowest() { return bit_cast<half_t>(binary_lowest); }
__host__ __device__ static constexpr half_t QuietNaN() { return bit_cast<half_t>(binary_qnan); }
};
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
template <>
struct NumericLimits<int4_t>
{
__host__ __device__ static constexpr int4_t Min() { return int4_t(-8); }
__host__ __device__ static constexpr int4_t Max() { return int4_t(7); }
__host__ __device__ static constexpr int4_t Lowest() { return int4_t(-8); }
};
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
template <>
struct NumericLimits<f8_t>
{
static constexpr uint8_t binary_min = 0x08; // 0b00001000
static constexpr uint8_t binary_max = 0x77; // 0b01110111
static constexpr uint8_t binary_lowest = 0xF7; // 0b11110111
static constexpr uint8_t binary_qnan = 0x80; // 0b10000000
__host__ __device__ static constexpr f8_t Min() { return bit_cast<f8_t>(binary_min); }
__host__ __device__ static constexpr f8_t Max() { return bit_cast<f8_t>(binary_max); }
__host__ __device__ static constexpr f8_t Lowest() { return bit_cast<f8_t>(binary_lowest); }
__host__ __device__ static constexpr f8_t QuietNaN() { return bit_cast<f8_t>(binary_qnan); }
};
#else
template <typename T>
struct NumericLimits
{
@@ -1151,6 +1308,7 @@ struct NumericLimits<bf8_t>
__host__ __device__ static constexpr bf8_t QuietNaN() { return bf8_t(binary_qnan); }
};
#endif
template <typename T>
struct NumericUtils

17
include/ck/utility/enable_if.hpp
View File

@@ -4,11 +4,26 @@
#pragma once
namespace ck {
#ifdef __HIPCC_RTC__
template <bool B, class T = void>
struct enable_if
{
};
template <class T>
struct enable_if<true, T>
{
using type = T;
};
template <bool B, class T = void>
using enable_if_t = typename enable_if<B, T>::type;
#else
template <bool B, typename T = void>
using enable_if = std::enable_if<B, T>;
template <bool B, typename T = void>
using enable_if_t = typename std::enable_if<B, T>::type;
#endif
} // namespace ck

4
include/ck/utility/env.hpp
View File

@@ -183,3 +183,7 @@ void UpdateEnvVar(EnvVar, const std::string_view& val)
}
} // namespace ck
// environment variable to enable logging:
// export CK_LOGGING=ON or CK_LOGGING=1 or CK_LOGGING=ENABLED
CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)

4
include/ck/utility/functional.hpp
View File

@@ -120,11 +120,11 @@ constexpr auto conditional_expr(X&& x, Y&& y)
{
if constexpr(predicate)
{
return std::forward<X>(x);
return ck::forward<X>(x);
}
else
{
return std::forward<Y>(y);
return ck::forward<Y>(y);
}
}

10
include/ck/utility/functional4.hpp
View File

@@ -21,7 +21,7 @@ struct unpack_impl<Sequence<Is...>>
template <typename F, typename X>
__host__ __device__ constexpr auto operator()(F&& f, X&& x) const
{
return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})...);
return ck::forward<F>(f)(ck::forward<X>(x).At(Number<Is>{})...);
}
};
@@ -35,8 +35,8 @@ struct unpack2_impl<Sequence<Is...>, Sequence<Js...>>
template <typename F, typename X, typename Y>
__host__ __device__ constexpr auto operator()(F&& f, X&& x, Y&& y) const
{
return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})...,
std::forward<Y>(y).At(Number<Js>{})...);
return ck::forward<F>(f)(ck::forward<X>(x).At(Number<Is>{})...,
ck::forward<Y>(y).At(Number<Js>{})...);
}
};
@@ -47,7 +47,7 @@ __host__ __device__ constexpr auto unpack(F&& f, X&& x)
{
using X_ = remove_reference_t<X>;
return detail::unpack_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type>{}(
std::forward<F>(f), std::forward<X>(x));
ck::forward<F>(f), ck::forward<X>(x));
}
// TODO: properly implement unpack that takes any number of containers
@@ -58,7 +58,7 @@ __host__ __device__ constexpr auto unpack2(F&& f, X&& x, Y&& y)
using Y_ = remove_reference_t<Y>;
return detail::unpack2_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type,
typename arithmetic_sequence_gen<0, Y_::Size(), 1>::type>{}(
std::forward<F>(f), std::forward<X>(x), std::forward<Y>(y));
ck::forward<F>(f), ck::forward<X>(x), ck::forward<Y>(y));
}
} // namespace ck

12
include/ck/utility/is_detected.hpp
View File

@@ -9,14 +9,14 @@ namespace detail {
template <class Default, class AlwaysVoid, template <class...> class Op, class... Args>
struct detector
{
using value_t = std::false_type;
using value_t = ck::false_type;
using type = Default;
};
template <class Default, template <class...> class Op, class... Args>
struct detector<Default, std::void_t<Op<Args...>>, Op, Args...>
struct detector<Default, ck::void_t<Op<Args...>>, Op, Args...>
{
using value_t = std::true_type;
using value_t = ck::true_type;
using type = Op<Args...>;
};
} // namespace detail
@@ -32,12 +32,12 @@ template <template <class...> class Op, class... Args>
using is_detected = typename detail::detector<nonesuch, void, Op, Args...>::value_t;
template <typename T>
using is_pack2_invocable_t = decltype(std::declval<T&>().is_pack2_invocable);
using is_pack2_invocable_t = decltype(ck::declval<T&>().is_pack2_invocable);
template <typename T>
using is_pack4_invocable_t = decltype(std::declval<T&>().is_pack4_invocable);
using is_pack4_invocable_t = decltype(ck::declval<T&>().is_pack4_invocable);
template <typename T>
using is_pack8_invocable_t = decltype(std::declval<T&>().is_pack8_invocable);
using is_pack8_invocable_t = decltype(ck::declval<T&>().is_pack8_invocable);
} // namespace ck

6
include/ck/utility/loop_scheduler.hpp
View File

@@ -1,8 +1,10 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ostream>
#pragma once
#ifndef __HIPCC_RTC__
#include <ostream>
#endif
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_adaptor.hpp"
@@ -26,6 +28,7 @@ constexpr LoopScheduler make_default_loop_scheduler()
} // namespace ck
#ifndef __HIPCC_RTC__
inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s)
{
switch(s)
@@ -36,3 +39,4 @@ inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s)
}
return os;
}
#endif

2
include/ck/utility/magic_division.hpp
View File

@@ -30,7 +30,7 @@ struct MagicDivision
// WARNING: magic division is only applicable for division inside this range.
// You should use the return value of CalculateMagicNumbers, if division is not inside this
// range. The "else" logic below is to quiet down run-time error.
if(divisor >= 1 && divisor <= INT32_MAX)
if(divisor >= 1 && divisor <= ck::NumericLimits<int32_t>::Max())
{
uint32_t shift = 0;
for(shift = 0; shift < 32; ++shift)

20
include/ck/utility/math_v2.hpp
View File

@@ -18,6 +18,7 @@ namespace math {
extern "C" __device__ float __ocml_native_recip_f32(float);
#endif
#ifndef __HIPCC_RTC__
// math functions for the host, some are implemented by calling C++ std functions
static inline __host__ float abs(float x) { return std::abs(x); };
@@ -457,6 +458,7 @@ inline __host__ double expm1<double>(double x)
{
return std::expm1(x);
}
#endif
// math functions for the HIP kernel, some are implemented by calling hip builtin functions
@@ -920,5 +922,23 @@ inline __device__ double expm1<double>(double x)
return expm1(x);
};
template <typename T>
inline __device__ T cos(T x)
{
return ck::type_convert<T>(cosf(ck::type_convert<float>(x)));
};
template <>
inline __device__ float cos<float>(float x)
{
return cosf(x);
};
template <>
inline __device__ double cos<double>(double x)
{
return cos(x);
};
} // namespace math
} // namespace ck

12
include/ck/utility/random_gen.hpp
View File

@@ -7,7 +7,7 @@ namespace ck {
// Pseudo random number generator
// version for fp32
template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<float, T>{}, bool> = false>
template <typename T, uint32_t seed_t, ck::enable_if_t<ck::is_same<float, T>{}, bool> = false>
__host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t)
{
uint32_t x = *(reinterpret_cast<uint32_t*>(&val));
@@ -23,7 +23,7 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed =
}
// version for fp16
template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<half_t, T>{}, bool> = false>
template <typename T, uint32_t seed_t, ck::enable_if_t<ck::is_same<half_t, T>{}, bool> = false>
__host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t)
{
uint16_t x = *(reinterpret_cast<uint16_t*>(&val));
@@ -40,12 +40,18 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed =
// return 0 if data is not fp16 or fp32
template <typename T,
uint32_t seed_t,
std::enable_if_t<!(std::is_same<float, T>{} || std::is_same<half_t, T>{}), bool> = false>
ck::enable_if_t<!(ck::is_same<float, T>{} || ck::is_same<half_t, T>{}), bool> = false>
__host__ __device__ uint32_t prand_generator(int id, T val, uint32_t seed = seed_t)
{
#ifdef __HIPCC_RTC__
static_cast<void>(id);
static_cast<void>(val);
static_cast<void>(seed);
#else
std::ignore = id;
std::ignore = val;
std::ignore = seed;
#endif
return 0;
}

4
include/ck/utility/sequence.hpp
View File

@@ -3,7 +3,9 @@
#pragma once
#ifndef __HIPCC_RTC__
#include <ostream>
#endif
#include "ck/utility/integral_constant.hpp"
#include "ck/utility/type.hpp"
@@ -900,6 +902,7 @@ using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
} // namespace ck
#ifndef __HIPCC_RTC__
template <ck::index_t... Is>
std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>)
{
@@ -910,3 +913,4 @@ std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>)
os << S::At(S::Size() - ck::Number<1>{}).value << "}";
return os;
}
#endif

14
include/ck/utility/tuple.hpp
View File

@@ -32,7 +32,7 @@ struct TupleElementKeyData
template <typename T,
typename enable_if<!is_same<remove_cvref_t<T>, TupleElementKeyData>::value,
bool>::type = false>
__host__ __device__ constexpr TupleElementKeyData(T&& v) : mData(std::forward<T>(v))
__host__ __device__ constexpr TupleElementKeyData(T&& v) : mData(ck::forward<T>(v))
{
}
@@ -67,7 +67,7 @@ get_tuple_element_data_reference(TupleElementKeyData<Key, Data>&& x)
template <typename Key, typename Data>
__host__ __device__ constexpr Data get_tuple_element_data(const TupleElementKeyData<Key, Data>& x)
{
return std::forward(x.mData);
return ck::forward(x.mData);
}
template <typename Indices, typename... Xs>
@@ -83,13 +83,13 @@ struct TupleImpl<Sequence<Is...>, Xs...> : TupleElementKeyData<TupleElementKey<I
!is_same<remove_cvref_t<Y>, TupleImpl>::value,
bool>::type = false>
__host__ __device__ constexpr TupleImpl(Y&& y)
: TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Y>(y))...
: TupleElementKeyData<TupleElementKey<Is>, Xs>(ck::forward<Y>(y))...
{
}
template <typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
__host__ __device__ constexpr TupleImpl(Ys&&... ys)
: TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Ys>(ys))...
: TupleElementKeyData<TupleElementKey<Is>, Xs>(ck::forward<Ys>(ys))...
{
static_assert(sizeof...(Is) == sizeof...(Xs) && sizeof...(Is) == sizeof...(Ys),
"wrong! inconsistent size");
@@ -123,14 +123,14 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
template <typename Y,
typename enable_if<sizeof...(Xs) == 1 && !is_same<remove_cvref_t<Y>, Tuple>::value,
bool>::type = false>
__host__ __device__ constexpr Tuple(Y&& y) : base(std::forward<Y>(y))
__host__ __device__ constexpr Tuple(Y&& y) : base(ck::forward<Y>(y))
{
}
template <typename... Ys,
typename enable_if<sizeof...(Ys) == sizeof...(Xs) && sizeof...(Ys) >= 2, bool>::type =
false>
__host__ __device__ constexpr Tuple(Ys&&... ys) : base(std::forward<Ys>(ys)...)
__host__ __device__ constexpr Tuple(Ys&&... ys) : base(ck::forward<Ys>(ys)...)
{
}
@@ -210,7 +210,7 @@ using tuple_element_t = typename tuple_element<I, TTuple>::type;
template <typename... Xs>
__host__ __device__ constexpr auto make_tuple(Xs&&... xs)
{
return Tuple<remove_cvref_t<Xs>...>(std::forward<Xs>(xs)...);
return Tuple<remove_cvref_t<Xs>...>(ck::forward<Xs>(xs)...);
}
// https://en.cppreference.com/w/cpp/utility/tuple/tie

6
include/ck/utility/tuple_helper.hpp
View File

@@ -29,7 +29,7 @@ __host__ __device__ constexpr auto concat_tuple_of_reference(const Tuple<X&...>&
const Tuple<Y&...>& ty)
{
return unpack2(
[&](auto&&... zs) { return Tuple<decltype(zs)...>{std::forward<decltype(zs)>(zs)...}; },
[&](auto&&... zs) { return Tuple<decltype(zs)...>{ck::forward<decltype(zs)>(zs)...}; },
tx,
ty);
}
@@ -38,7 +38,7 @@ template <typename... X, typename... Y>
__host__ __device__ constexpr auto concat_tuple(const Tuple<X...>& tx, const Tuple<Y...>& ty)
{
return unpack2(
[&](auto... zs) { return Tuple<decltype(zs)...>{std::forward<decltype(zs)>(zs)...}; },
[&](auto... zs) { return Tuple<decltype(zs)...>{ck::forward<decltype(zs)>(zs)...}; },
tx,
ty);
}
@@ -157,6 +157,7 @@ __host__ __device__ constexpr auto TupleReduce(F&& f, const Tuple<Ts...>& tuple)
}
}
#ifndef __HIPCC_RTC__
template <typename T>
using is_tuple = decltype(std::declval<T&>().IsTuple());
@@ -165,6 +166,7 @@ __host__ __device__ constexpr auto IsNestedTuple(const Tuple<Ts...>&)
{
return (is_detected<is_tuple, Ts>::value || ...);
}
#endif
template <index_t depth = 0, typename T>
__host__ __device__ constexpr auto TupleDepth(const T&)

162
include/ck/utility/type.hpp
View File

@@ -8,6 +8,158 @@
#include "ck/utility/enable_if.hpp"
namespace ck {
#ifdef __HIPCC_RTC__
template <bool B>
using bool_constant = integral_constant<bool, B>;
using true_type = bool_constant<true>;
using false_type = bool_constant<false>;
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAIT1(name) \
template <class T> \
struct name : bool_constant<__##name(T)> \
{ \
}
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAIT2(name) \
template <class T, class U> \
struct name : bool_constant<__##name(T, U)> \
{ \
}
// NOLINTNEXTLINE
#define CK_BUILTIN_TYPE_TRAITN(name) \
template <class... Ts> \
struct name : bool_constant<__##name(Ts...)> \
{ \
}
CK_BUILTIN_TYPE_TRAIT1(is_class);
CK_BUILTIN_TYPE_TRAIT1(is_pointer);
CK_BUILTIN_TYPE_TRAIT1(is_reference);
CK_BUILTIN_TYPE_TRAIT1(is_trivially_copyable);
CK_BUILTIN_TYPE_TRAIT1(is_unsigned);
CK_BUILTIN_TYPE_TRAIT2(is_base_of);
template <class T>
struct remove_cv
{
using type = T;
};
template <class T>
struct remove_cv<const T> : remove_cv<T>
{
};
template <class T>
struct remove_cv<volatile T> : remove_cv<T>
{
};
template <class T>
struct remove_reference
{
typedef T type;
};
template <class T>
struct remove_reference<T&>
{
typedef T type;
};
template <class T>
struct remove_reference<T&&>
{
typedef T type;
};
template <class T>
struct remove_pointer
{
typedef T type;
};
template <class T>
struct remove_pointer<T*>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* const>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* volatile>
{
typedef T type;
};
template <class T>
struct remove_pointer<T* const volatile>
{
typedef T type;
};
template <typename T>
constexpr T&& forward(typename remove_reference<T>::type& t_) noexcept
{
return static_cast<T&&>(t_);
}
template <typename T>
constexpr T&& forward(typename remove_reference<T>::type&& t_) noexcept
{
return static_cast<T&&>(t_);
}
// TODO
template<class T> struct is_const : false_type {};
template<class T> struct is_const<const T> : true_type {};
template< class T >
inline constexpr bool is_const_v = is_const<T>::value;
template< class T >
inline constexpr bool is_reference_v = is_reference<T>::value;
template<class T> struct remove_const { typedef T type; };
template<class T> struct remove_const<const T> { typedef T type; };
template< class T >
using remove_const_t = typename remove_const<T>::type;
template< class T >
inline constexpr bool is_class_v = is_class<T>::value;
template< class T >
inline constexpr bool is_trivially_copyable_v = is_trivially_copyable<T>::value;
template< class... >
using void_t = void;
using __hip::declval;
#else
#include <utility>
#include <type_traits>
using std::forward;
using std::is_base_of;
using std::is_class;
using std::is_pointer;
using std::is_reference;
using std::is_trivially_copyable;
using std::is_unsigned;
using std::remove_cv;
using std::remove_pointer;
using std::remove_reference;
using std::is_const_v;
using std::is_reference_v;
using std::remove_const_t;
using std::is_class_v;
using std::is_trivially_copyable_v;
using std::void_t;
using std::false_type;
using std::true_type;
using std::declval;
#endif
template <typename X, typename Y>
struct is_same : public integral_constant<bool, false>
@@ -23,19 +175,19 @@ template <typename X, typename Y>
inline constexpr bool is_same_v = is_same<X, Y>::value;
template <typename T>
using remove_reference_t = typename std::remove_reference<T>::type;
using remove_reference_t = typename remove_reference<T>::type;
template <typename T>
using remove_cv_t = typename std::remove_cv<T>::type;
using remove_cv_t = typename remove_cv<T>::type;
template <typename T>
using remove_cvref_t = remove_cv_t<std::remove_reference_t<T>>;
using remove_cvref_t = remove_cv_t<remove_reference_t<T>>;
template <typename T>
using remove_pointer_t = typename std::remove_pointer<T>::type;
using remove_pointer_t = typename remove_pointer<T>::type;
template <typename T>
inline constexpr bool is_pointer_v = std::is_pointer<T>::value;
inline constexpr bool is_pointer_v = is_pointer<T>::value;
template <typename Y, typename X, typename enable_if<sizeof(X) == sizeof(Y), bool>::type = false>
__host__ __device__ constexpr Y bit_cast(const X& x)

24
include/ck/utility/type_convert.hpp
View File

@@ -17,10 +17,10 @@ namespace ck {
// Convert X to Y, both X and Y are non-const data types.
template <typename Y,
typename X,
std::enable_if_t<!(std::is_const_v<Y> || std::is_const_v<X>), bool> = false>
ck::enable_if_t<!(ck::is_const_v<Y> || ck::is_const_v<X>), bool> = false>
__host__ __device__ constexpr Y type_convert(X x)
{
static_assert(!std::is_reference_v<Y> && !std::is_reference_v<X>);
static_assert(!ck::is_reference_v<Y> && !ck::is_reference_v<X>);
return static_cast<Y>(x);
}
@@ -28,13 +28,13 @@ __host__ __device__ constexpr Y type_convert(X x)
// Convert X to Y, either X or Y is a const data type.
template <typename Y,
typename X,
std::enable_if_t<std::is_const_v<Y> || std::is_const_v<X>, bool> = false>
ck::enable_if_t<ck::is_const_v<Y> || ck::is_const_v<X>, bool> = false>
__host__ __device__ constexpr Y type_convert(X x)
{
static_assert(!std::is_reference_v<Y> && !std::is_reference_v<X>);
static_assert(!ck::is_reference_v<Y> && !ck::is_reference_v<X>);
using NonConstY = std::remove_const_t<Y>;
using NonConstX = std::remove_const_t<X>;
using NonConstY = ck::remove_const_t<Y>;
using NonConstX = ck::remove_const_t<X>;
return static_cast<Y>(type_convert<NonConstY, NonConstX>(x));
}
@@ -104,7 +104,7 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
template <typename Y, typename X>
__host__ __device__ constexpr Y type_convert_sp(X x)
{
static_assert(!std::is_reference_v<Y> && !std::is_reference_v<X>);
static_assert(!ck::is_reference_v<Y> && !ck::is_reference_v<X>);
return static_cast<Y>(x);
}
@@ -166,7 +166,7 @@ template <>
inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
{
constexpr int seed = 1254739;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<long_index_t>(&x), x);
#if defined(__gfx94__)
union
{
@@ -206,7 +206,7 @@ inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 1254739;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<long_index_t>(&x), x);
return utils::
cast_to_f8<half_t, f8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
@@ -218,7 +218,7 @@ template <>
inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, float>(float x)
{
constexpr int seed = 1254739;
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
uint32_t rng = prand_generator<float, seed>(reinterpret_cast<long_index_t>(&x), x);
#if defined(__gfx94__)
union
{
@@ -258,7 +258,7 @@ inline __host__ __device__ bf8_t f8_convert_sr<bf8_t, half_t>(half_t x)
constexpr bool clip = true;
constexpr f8_rounding_mode rm = f8_rounding_mode::stochastic;
constexpr int seed = 1254739;
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<long_index_t>(&x), x);
return utils::
cast_to_f8<half_t, bf8_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
x, rng);
@@ -501,6 +501,7 @@ inline __host__ __device__ half_t type_convert<half_t, bf8_t>(bf8_t x)
#endif
}
#ifndef __HIPCC_RTC__
template <typename Y, typename X, std::size_t NumElems>
inline __host__ __device__ void array_convert(std::array<Y, NumElems>& y,
const std::array<X, NumElems>& x)
@@ -510,6 +511,7 @@ inline __host__ __device__ void array_convert(std::array<Y, NumElems>& y,
y[i] = type_convert<Y>(x[i]);
}
}
#endif
template <typename Y, typename X, index_t NumElems>
inline __host__ __device__ void array_convert(Array<Y, NumElems>& y, const Array<X, NumElems>& x)