ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 02:02:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add support for mixed-precision f16bf16_int8 gemm (#1127)

Browse Source 
[ROCm/composable_kernel commit: ba86eadce5]
This commit is contained in:
jakpiase
2024-02-07 15:54:13 +01:00
committed by GitHub
parent f59cd842c8
commit b41cf51951
33 changed files with 2424 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -75,6 +75,15 @@ struct Add
y = ck::type_convert<bhalf_t>(y_tmp);
}
template <>
__host__ __device__ constexpr void
operator()<bhalf_t>(bhalf_t& y, const float& x0, const bhalf_t& x1) const
{
const float x2_tmp = ck::type_convert<float>(x1);
const float y_tmp = x0 + x2_tmp;
y = ck::type_convert<bhalf_t>(y_tmp);
}
template <>
__host__ __device__ constexpr void
operator()<int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
@@ -264,6 +273,14 @@ struct AddRelu
y = a > 0.0f ? a : 0.0f;
};
template <>
__host__ __device__ constexpr void
operator()<bhalf_t, float, bhalf_t>(bhalf_t& y, const float& x0, const bhalf_t& x1) const
{
const float a = x0 + type_convert<float>(x1);
y = a > type_convert<bhalf_t>(0.0f) ? a : type_convert<bhalf_t>(0.0f);
};
template <>
__host__ __device__ constexpr void
operator()<int, int, int8_t>(int& y, const int& x0, const int8_t& x1) const
@@ -354,6 +371,70 @@ struct AddFastGelu
e = type_convert<half_t>(x1_f);
}
template <>
__host__ __device__ constexpr void
operator()<bhalf_t, float, bhalf_t>(bhalf_t& e, const float& c, const bhalf_t& d) const
{
const float x0_f = c + type_convert<float>(d);
float x1_f = 0;
FastGelu{}.template operator()<float, float>(x1_f, x0_f);
e = type_convert<bhalf_t>(x1_f);
}
};
// E = Silu(C + D)
struct AddSilu
{
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void
operator()<float, float, float>(float& e, const float& c, const float& d) const
{
const float x = c + d;
Silu{}.template operator()<float>(e, x);
}
template <>
__host__ __device__ constexpr void
operator()<half_t, half_t, half_t>(half_t& e, const half_t& c, const half_t& d) const
{
const half_t x = c + d;
Silu{}.template operator()<half_t>(e, x);
}
template <>
__host__ __device__ constexpr void
operator()<half_t, float, half_t>(half_t& e, const float& c, const half_t& d) const
{
const float x0_f = c + d;
float x1_f = 0;
Silu{}.template operator()<float>(x1_f, x0_f);
e = type_convert<half_t>(x1_f);
}
template <>
__host__ __device__ constexpr void
operator()<bhalf_t, float, bhalf_t>(bhalf_t& e, const float& c, const bhalf_t& d) const
{
const float x0_f = c + type_convert<float>(d);
float x1_f = 0;
Silu{}.template operator()<float>(x1_f, x0_f);
e = type_convert<bhalf_t>(x1_f);
}
};
} // namespace element_wise

21
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -156,6 +156,12 @@ struct PassThrough
y = type_convert<half_t>(x);
}
template <>
__host__ __device__ void operator()<bhalf_t, int8_t>(bhalf_t& y, const int8_t& x) const
{
y = type_convert<bhalf_t>(x);
}
template <>
__host__ __device__ void operator()<int8_t, int32_t>(int8_t& y, const int32_t& x) const
{
@@ -551,6 +557,19 @@ struct Sigmoid
};
};
struct Silu
{
template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const
{
static_assert(is_same_v<T, float> || is_same_v<T, double> || is_same_v<T, ck::half_t> ||
is_same_v<T, int8_t> || is_same_v<T, int32_t>,
"Data type is not supported by this operation!");
constexpr T one = type_convert<T>(1);
y = x * (one / (one + ck::math::exp(-x)));
};
};
struct TanH
{
template <typename T>

4
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -98,6 +98,8 @@ using Scale = ck::tensor_operation::element_wise::Scale;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
using AddRelu = ck::tensor_operation::element_wise::AddRelu;
using AddSilu = ck::tensor_operation::element_wise::AddSilu;
using AddReluAdd = ck::tensor_operation::element_wise::AddReluAdd;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;

114
library/include/ck/library/tensor_operation_instance/gpu/gemm_add.hpp Normal file
View File

@@ -0,0 +1,114 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Add>>>&);
void add_device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Add>>>&);
// GEMM + Add +
template <typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout,
typename ADataType,
typename BDataType,
typename D0DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
Add>>
{
using DeviceOp = DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
Add>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_ENABLE_INT8) && defined(CK_ENABLE_FP16)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(op_ptrs);
}
}
#endif
#if defined(CK_ENABLE_INT8) && defined(CK_ENABLE_BF16)
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, ck::bhalf_t> && is_same_v<EDataType, ck::bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

54
library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -68,6 +68,32 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_inst
PassThrough,
AddFastGelu>>>&);
void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
AddFastGelu>>>&);
void add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
AddFastGelu>>>&);
// GEMM + Add + FastGelu
template <typename ALayout,
typename BLayout,
@@ -106,6 +132,32 @@ struct DeviceOperationInstanceFactory<
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_ENABLE_INT8) && defined(CK_ENABLE_FP16)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
op_ptrs);
}
}
#endif
#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, bhalf_t> && is_same_v<EDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
op_ptrs);
}
}
#endif
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
{

116
library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu.hpp Normal file
View File

@@ -0,0 +1,116 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
AddRelu>>>&);
void add_device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
AddRelu>>>&);
// GEMM + Add + Relu
template <typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout,
typename ADataType,
typename BDataType,
typename D0DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddRelu>>
{
using DeviceOp = DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddRelu>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_ENABLE_INT8) && defined(CK_ENABLE_FP16)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
op_ptrs);
}
}
#endif
#if defined(CK_ENABLE_INT8) && defined(CK_ENABLE_BF16)
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, ck::bhalf_t> && is_same_v<EDataType, ck::bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

116
library/include/ck/library/tensor_operation_instance/gpu/gemm_add_silu.hpp Normal file
View File

@@ -0,0 +1,116 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
AddSilu>>>&);
void add_device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
AddSilu>>>&);
// GEMM + Add + Silu
template <typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout,
typename ADataType,
typename BDataType,
typename D0DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddSilu>>
{
using DeviceOp = DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
PassThrough,
PassThrough,
AddSilu>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if defined(CK_ENABLE_INT8) && defined(CK_ENABLE_FP16)
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
op_ptrs);
}
}
#endif
#if defined(CK_ENABLE_INT8) && defined(CK_ENABLE_BF16)
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, int8_t> &&
is_same_v<D0DataType, ck::bhalf_t> && is_same_v<EDataType, ck::bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
{
add_device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

4
library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt Normal file
View File

@@ -0,0 +1,4 @@
add_instance_library(device_gemm_add_instance
device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
)

69
library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,69 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
Add>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances, device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

69
library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,69 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, Add, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
Add>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances, device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

2
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
View File

@@ -1,6 +1,8 @@
add_instance_library(device_gemm_add_fastgelu_instance
device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
)

73
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,73 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances =
std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
AddFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances,
device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

72
library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,72 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances =
std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
AddFastGelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

4
library/src/tensor_operation_instance/gpu/gemm_add_relu/CMakeLists.txt Normal file
View File

@@ -0,0 +1,4 @@
add_instance_library(device_gemm_add_relu_instance
device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
)

71
library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,71 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances =
std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
AddRelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances, device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

70
library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,70 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddRelu, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
AddRelu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances, device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

4
library/src/tensor_operation_instance/gpu/gemm_add_silu/CMakeLists.txt Normal file
View File

@@ -0,0 +1,4 @@
add_instance_library(device_gemm_add_silu_instance
device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
)

71
library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,71 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances =
std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, BF16, I8, F32, F32,BF16_Tuple, BF16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
BF16,
I8,
BF16_Tuple,
BF16,
PassThrough,
PassThrough,
AddSilu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances, device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

70
library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp Normal file
View File

@@ -0,0 +1,70 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/utility/sequence.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// e = elementwise((a * b), d0, d1)
// outout: e[m, n]
// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
using device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 1>
// clang-format on
>;
using device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances = std::tuple<
// clang-format off
// M/N/K padding
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 256, 16, 128, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, I8, F32, F32, F16_Tuple, F16, PassThrough, PassThrough, AddSilu, GemmMNKPadding, 1, 64, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 1, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, PipelineVersion::v1>
// clang-format on
>;
void add_device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Tuple,
Row,
F16,
I8,
F16_Tuple,
F16,
PassThrough,
PassThrough,
AddSilu>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_generic_instances{});
add_device_operation_instances(
instances, device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

232
profiler/include/profiler/profile_gemm_add_impl.hpp Normal file
View File

@@ -0,0 +1,232 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.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_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Add>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck

232
profiler/include/profiler/profile_gemm_add_relu_impl.hpp Normal file
View File

@@ -0,0 +1,232 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.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_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_relu_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddRelu = ck::tensor_operation::element_wise::AddRelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddRelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddRelu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck

232
profiler/include/profiler/profile_gemm_add_silu_impl.hpp Normal file
View File

@@ -0,0 +1,232 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add_silu.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.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_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_silu_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddRelu = ck::tensor_operation::element_wise::AddSilu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddRelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddSilu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck

6
profiler/src/CMakeLists.txt
View File

@@ -43,7 +43,10 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
@@ -109,7 +112,10 @@ if(DL_KERNELS)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)

139
profiler/src/profile_gemm_add.cpp Normal file
View File

@@ -0,0 +1,139 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_add_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_add"
#define OP_DESC "GEMM+Add"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F16_INT8_F16_F16, // 0
BF16_INT8_BF16_BF16, // 1
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: f16&i8 1: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = ReLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = ReLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = ReLU(A[k, m] * B[k, n] + D0[m, n]);\n");
printf(" 3: E[m, n] = ReLU(A[k, m] * B[n, k] + D0[m, n]))\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD0 = std::stoi(argv[13]);
const int StrideE = std::stoi(argv[14]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
// using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d0_type,
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto e_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using D0DataType = decltype(d0_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using ELayout = decltype(e_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add);

17
profiler/src/profile_gemm_add_fastgelu.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
@@ -12,6 +12,9 @@
#define OP_NAME "gemm_add_fastgelu"
#define OP_DESC "GEMM+Add+FastGeLU"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add_fastgelu(int argc, char* argv[])
{
enum struct MatrixLayout
@@ -28,13 +31,15 @@ int profile_gemm_add_fastgelu(int argc, char* argv[])
F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8, // 3
F16_INT8_F16_F16, // 4
BF16_INT8_BF16_BF16, // 5
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f16&i8 5: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n]);\n");
@@ -135,6 +140,14 @@ int profile_gemm_add_fastgelu(int argc, char* argv[])
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;

139
profiler/src/profile_gemm_add_relu.cpp Normal file
View File

@@ -0,0 +1,139 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_add_relu_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_add_relu"
#define OP_DESC "GEMM+Add+ReLU"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add_relu(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F16_INT8_F16_F16, // 0
BF16_INT8_BF16_BF16, // 1
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: f16&i8 1: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = ReLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = ReLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = ReLU(A[k, m] * B[k, n] + D0[m, n]);\n");
printf(" 3: E[m, n] = ReLU(A[k, m] * B[n, k] + D0[m, n]))\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD0 = std::stoi(argv[13]);
const int StrideE = std::stoi(argv[14]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
// using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d0_type,
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto e_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using D0DataType = decltype(d0_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using ELayout = decltype(e_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_relu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_relu);

139
profiler/src/profile_gemm_add_silu.cpp Normal file
View File

@@ -0,0 +1,139 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_add_silu_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_add_silu"
#define OP_DESC "GEMM+Add+SiLU"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add_silu(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F16_INT8_F16_F16, // 0
BF16_INT8_BF16_BF16, // 1
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: f16&i8 1: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = ReLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = ReLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = ReLU(A[k, m] * B[k, n] + D0[m, n]);\n");
printf(" 3: E[m, n] = ReLU(A[k, m] * B[n, k] + D0[m, n]))\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD0 = std::stoi(argv[13]);
const int StrideE = std::stoi(argv[14]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
// using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d0_type,
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto e_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using D0DataType = decltype(d0_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using ELayout = decltype(e_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_silu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_silu);

1
test/CMakeLists.txt
View File

@@ -122,6 +122,7 @@ add_subdirectory(space_filling_curve)
add_subdirectory(conv_util)
add_subdirectory(reference_conv_fwd)
add_subdirectory(gemm)
add_subdirectory(gemm_add)
add_subdirectory(gemm_layernorm)
add_subdirectory(gemm_split_k)
add_subdirectory(gemm_reduce)

11
test/gemm_add/CMakeLists.txt Normal file
View File

@@ -0,0 +1,11 @@
add_gtest_executable(test_gemm_add test_gemm_add.hpp)
target_link_libraries(test_gemm_add PRIVATE utility device_gemm_add_instance)
add_gtest_executable(test_gemm_add_relu test_gemm_add_relu.cpp)
target_link_libraries(test_gemm_add_relu PRIVATE utility device_gemm_add_instance device_gemm_add_relu_instance)
add_gtest_executable(test_gemm_add_silu test_gemm_add_silu.cpp)
target_link_libraries(test_gemm_add_silu PRIVATE utility device_gemm_add_instance device_gemm_add_silu_instance)
add_gtest_executable(test_gemm_add_fastgelu test_gemm_add_fastgelu.cpp)
target_link_libraries(test_gemm_add_fastgelu PRIVATE utility device_gemm_add_instance device_gemm_add_fastgelu_instance)

72
test/gemm_add/test_gemm_add.hpp Normal file
View File

@@ -0,0 +1,72 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_impl.hpp"
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using I8 = int8_t;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <typename Tuple>
class TestGemmAdd : public ::testing::Test
{
protected:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddImpl = ck::profiler::profile_gemm_add_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
virtual decltype(ProfileGemmAddImpl) GetImpl() { return ProfileGemmAddImpl; }
void Run()
{
std::vector<std::vector<ck::index_t>> lengths = {
{16, 32, 64}, {2048, 4096, 8192}, {2048, 1024, 16}};
bool all_success = true;
for(auto length : lengths)
{
int M = length[0];
int N = length[1];
int K = length[2];
int StrideA = ck::is_same_v<ALayout, Row> ? K : M;
int StrideB = ck::is_same_v<BLayout, Row> ? N : K;
int StrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
int StrideE = ck::is_same_v<ELayout, Row> ? N : M;
all_success =
all_success &
GetImpl()(true, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideE);
}
EXPECT_TRUE(all_success);
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAdd, KernelTypes);
TYPED_TEST(TestGemmAdd, Test_BF16FP16_INT8) { this->Run(); }

41
test/gemm_add/test_gemm_add_fastgelu.cpp Normal file
View File

@@ -0,0 +1,41 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_fastgelu_impl.hpp"
#include "test_gemm_add.hpp"
template <typename Tuple>
class TestGemmAddFastgelu : public TestGemmAdd<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddFastgeluImpl =
ck::profiler::profile_gemm_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; }
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAddFastgelu, KernelTypes);
TYPED_TEST(TestGemmAddFastgelu, Test_BF16FP16) { this->Run(); }

41
test/gemm_add/test_gemm_add_relu.cpp Normal file
View File

@@ -0,0 +1,41 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_relu_impl.hpp"
#include "test_gemm_add.hpp"
template <typename Tuple>
class TestGemmAddRelu : public TestGemmAdd<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddReluImpl =
ck::profiler::profile_gemm_add_relu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddReluImpl) GetImpl() override { return ProfileGemmAddReluImpl; }
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAddRelu, KernelTypes);
TYPED_TEST(TestGemmAddRelu, Test_BF16FP16_INT8) { this->Run(); }

41
test/gemm_add/test_gemm_add_silu.cpp Normal file
View File

@@ -0,0 +1,41 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_silu_impl.hpp"
#include "test_gemm_add.hpp"
template <typename Tuple>
class TestGemmAddSilu : public TestGemmAdd<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddSiluImpl =
ck::profiler::profile_gemm_add_silu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddSiluImpl) GetImpl() override { return ProfileGemmAddSiluImpl; }
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAddSilu, KernelTypes);
TYPED_TEST(TestGemmAddSilu, Test_BF16FP16_INT8) { this->Run(); }