Skip to content

GitLab

Unverified Commit 1d8e4ec2 authored by Adam Osewski's avatar Adam Osewski Committed by GitHub
Browse files

Jing's contribution: prototype of mixed precision gemm FP16/BF16xint4 GEMM (#1762) 



* add a prototype of int4

* clean

* debug

* clean

* clean

* move packed into dynamic_buffer

* fixed coord reset

* add fast pki4 to half conversion

* fix

* fixed reference and host_tensor

* fixed tensor init

* format

* debug i4_to_f16_convert

* format

* fixed splitk

* weight permute

* add b tile permute

* clean

* weight permute with splitki

* format

* improve weight layout

* add and_or_b32

* fixed splitk crush

* add permute switch as a template

* recover v3r1

* clean

* failure with intrawave v2

* fixed

* fixed

* add ckProfiler

* add bfp16 support

* add bf16 example

* fixed int4 to bhalf_t conversion

* format

* fixed int4 to bf16 conversion

* clean

* add instances for mem

* clean

* fixed host tensor size

* fixed

* debug

* fixed

* add pk_i4_t as a struct

* fix

* Update example/01_gemm/gemm_xdl_bf16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update example/01_gemm/gemm_xdl_bf16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update example/01_gemm/gemm_xdl_bf16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* revert

* Update example/01_gemm/gemm_xdl_bf16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update example/01_gemm/gemm_xdl_fp16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update example/01_gemm/gemm_xdl_fp16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update example/01_gemm/gemm_xdl_fp16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update example/01_gemm/gemm_xdl_fp16_pk_i4_v3.cpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* fixed comments

* revert

* clean

* revert

* revert

* fixed

* Update CMakeLists.txt

* Update script/cmake-ck-dev.sh
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* Update CMakeLists.txt
Co-authored-by: default avatarAdam Osewski <19374865+aosewski@users.noreply.github.com>

* fixed

* fixed

* fixed

* revert

* revert

* add comments

* format

* fixed assert

* fixed

* Fix I4 define in ckProfiler

* Fixed example_gemm_xdl_bf16_pk_i4_v3 test failed issue

---------
Co-authored-by: default avatarJing Zhang <jizhan@fb.com>
Co-authored-by: default avatarzjing14 <zhangjing14@gmail.com>
Co-authored-by: default avatarmtgu0705 <mtgu@amd.com>
parent 159fa319
Hide whitespace changes
Inline Side-by-side
Showing with 458 additions and 16 deletions
include/ck/utility/amd_buffer_addressing.hpp
View file @ 1d8e4ec2
......@@ -429,7 +429,8 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w
(is_same<T, f8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, bf8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, uint8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
(is_same<T, uint8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, pk_i4_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
"wrong! not implemented");
using r_t = typename vector_type<T, N>::type;
......
include/ck/utility/amd_inline_asm.hpp
View file @ 1d8e4ec2
......@@ -11,6 +11,27 @@
namespace ck {
inline __device__ int amd_assembly_and_or_b32(int a, int b, int d)
{
int c;
asm volatile("v_and_or_b32 %0, %1, %2, %3" : "=v"(c) : "v"(a), "v"(b), "v"(d));
return c;
}
inline __device__ half2_t amd_assembly_pk_fma_f16(half2_t a, half2_t b, half2_t c)
{
half2_t d;
asm volatile("v_pk_fma_f16 %0, %1, %2, %3;\n" : "=v"(d) : "v"(a), "v"(b), "v"(c));
return d;
}
inline __device__ half2_t amd_assembly_pk_add_f16(half2_t a, half2_t b)
{
half2_t c;
asm volatile("v_pk_add_f16 %0, %1, %2;\n" : "=v"(c) : "v"(a), "v"(b));
return c;
}
// c0 += inner_product(a, b0)
// c1 += inner_product(a, b1)
__device__ void amd_assembly_outer_product_1x2(float a, float b0, float b1, float& c0, float& c1)
......
include/ck/utility/data_type.hpp
View file @ 1d8e4ec2
......@@ -12,6 +12,15 @@ using bhalf_t = ushort;
using half_t = _Float16;
using int4_t = _BitInt(4);
// custom data type - pack int4 data
struct pk_i4_t
{
using type = int8_t;
type data;
__host__ __device__ constexpr pk_i4_t() : data{type{}} {}
__host__ __device__ constexpr pk_i4_t(type init) : data{init} {}
};
inline constexpr auto next_pow2(uint32_t x)
{
// Precondition: x > 1.
......@@ -165,6 +174,13 @@ struct scalar_type<int4_t>
};
#endif
template <>
struct scalar_type<pk_i4_t>
{
using type = pk_i4_t;
static constexpr index_t vector_size = 1;
};
template <>
struct scalar_type<f8_fnuz_t>
{
......@@ -1044,6 +1060,12 @@ struct nnvb_data_t_selector<bf8_ocp_t>
using type = bf8_ocp_t::data_type;
};
template <>
struct nnvb_data_t_selector<pk_i4_t>
{
using type = pk_i4_t::type;
};
template <typename T, index_t N>
struct non_native_vector_base<
T,
......@@ -1163,6 +1185,14 @@ struct scalar_type<non_native_vector_base<bf8_ocp_t, N>>
static constexpr index_t vector_size = N;
};
template <index_t N>
struct scalar_type<non_native_vector_base<pk_i4_t, N>>
{
using type = typename non_native_vector_base<pk_i4_t, N>::data_t;
static constexpr index_t vector_size = N;
};
// non-native vector_type implementation
template <typename T>
struct vector_type<T, 1, typename std::enable_if_t<!is_native_type<T>()>>
......@@ -1871,6 +1901,11 @@ using uint8x16_t = typename vector_type<uint8_t, 16>::type;
using uint8x32_t = typename vector_type<uint8_t, 32>::type;
using uint8x64_t = typename vector_type<uint8_t, 64>::type;
// pack int4
using pk_i4x2_t = typename vector_type<pk_i4_t, 2>::type;
using pk_i4x4_t = typename vector_type<pk_i4_t, 4>::type;
using pk_i4x8_t = typename vector_type<pk_i4_t, 8>::type;
template <typename T>
struct NumericLimits
{
......
include/ck/utility/dynamic_buffer.hpp
View file @ 1d8e4ec2
......@@ -54,7 +54,8 @@ struct DynamicBuffer
template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value,
typename scalar_type<remove_cvref_t<T>>::type>::value ||
!is_native_type<X>(),
bool>::type = false>
__host__ __device__ constexpr auto Get(index_t i, bool is_valid_element) const
{
......@@ -195,7 +196,8 @@ struct DynamicBuffer
template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value,
typename scalar_type<remove_cvref_t<T>>::type>::value ||
!is_native_type<X>(),
bool>::type = false>
__host__ __device__ void Set(index_t i, bool is_valid_element, const X& x)
{
......
include/ck/utility/static_buffer.hpp
View file @ 1d8e4ec2
......@@ -116,7 +116,8 @@ struct StaticBufferTupleOfVector
// i is offset of S, not X. i should be aligned to X
template <typename X,
index_t I,
typename enable_if<has_same_scalar_type<S, X>::value, bool>::type = false>
typename enable_if<has_same_scalar_type<S, X>::value || !is_native_type<S>(),
bool>::type = false>
__host__ __device__ constexpr auto GetAsType(Number<I> i) const
{
constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};
......@@ -134,7 +135,8 @@ struct StaticBufferTupleOfVector
// i is offset of S, not X. i should be aligned to X
template <typename X,
index_t I,
typename enable_if<has_same_scalar_type<S, X>::value, bool>::type = false>
typename enable_if<has_same_scalar_type<S, X>::value || !is_native_type<S>(),
bool>::type = false>
__host__ __device__ constexpr void SetAsType(Number<I> i, X x)
{
constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ 1d8e4ec2
......@@ -74,6 +74,17 @@ struct ReferenceGemm : public device::BaseOperator
{
ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_m_k_(m, k));
}
else if constexpr(is_same_v<ADataType, pk_i4_t>)
{
uint8_t i4x2 = arg.a_m_k_(m, k).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2 >> 0) & 0xf;
else
i4 = (i4x2 >> 4) & 0xf;
i4 = i4 - 8;
v_a = type_convert<ComputeTypeA>(i4);
}
else
{
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
......@@ -84,6 +95,17 @@ struct ReferenceGemm : public device::BaseOperator
{
ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_k_n_(k, n));
}
else if constexpr(is_same_v<BDataType, pk_i4_t>)
{
uint8_t i4x2 = arg.b_k_n_(k, n).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2 >> 0) & 0xf;
else
i4 = (i4x2 >> 4) & 0xf;
i4 = i4 - 8;
v_b = type_convert<ComputeTypeB>(i4);
}
else
{
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
......
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ 1d8e4ec2
......@@ -22,6 +22,7 @@ using I8 = int8_t;
using I32 = int32_t;
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
using I4 = ck::pk_i4_t;
using Empty_Tuple = ck::Tuple<>;
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp
View file @ 1d8e4ec2
......@@ -166,11 +166,22 @@ void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instances
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, I4, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, I4, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -810,6 +821,28 @@ struct DeviceOperationInstanceFactory<
}
}
#endif
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, pk_i4_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
}
}
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, pk_i4_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
}
}
return op_ptrs;
}
};
......
library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt
View file @ 1d8e4ec2
......@@ -97,6 +97,9 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f16_f8_f16/device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
device_gemm_xdl_universal_f16_i4_f16/device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_bf16_i4_bf16/device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_mem_v2_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp
device_gemm_xdl_universal_f8_f16_f16/device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp
......
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_i4_bf16/device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn.hpp 0 → 100644
View file @ 1d8e4ec2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using I4 = pk_i4_t;
using BF16 = bhalf_t;
using F32 = float;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
#if 0
template <GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_comp_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| 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| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
#endif
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_mem_instances =
std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| 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| Block-wiseGemm| Block-wiseGemm| ACompType| BCompType| APermute| BPermute|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| | | | |
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| | | | |
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 8, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 32, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 32, 128, 8, 32, 32, 32, 2, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 128, 8, 16, 16, 16, 4, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 32, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 128, 8, 16, 16, 16, 2, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 32, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 128, 8, 32, 16, 16, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 64, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 128, 128, 8, 32, 32, 32, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 256, 128, 8, 32, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, bhalf_t, bhalf_t, false, true>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_i4_bf16/device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_mem_v2_default_instance.cpp 0 → 100644
View file @ 1d8e4ec2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, I4, BF16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_bf16_i4_bf16_mk_nk_mn_mem_instances<Interwave, GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_i4_f16/device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn.hpp 0 → 100644
View file @ 1d8e4ec2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using I4 = pk_i4_t;
using F16 = half_t;
using F32 = float;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
#if 0
template <GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_comp_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| 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| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
#endif
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_mem_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle| A| B| C| GEMM| 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| Block-wiseGemm| Block-wiseGemm| ACompType| BCompType| APermute| BPermute|
//#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| | | | |
//#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| | | | |
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 8, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 32, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 32, 128, 8, 32, 32, 32, 2, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 128, 8, 16, 16, 16, 4, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 32, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 128, 8, 16, 16, 16, 2, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 128, 8, 16, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 128, 8, 32, 16, 16, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 128, 8, 32, 16, 16, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 64, 128, 8, 32, 32, 32, 1, 1, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 128, 128, 8, 32, 32, 32, 1, 2, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 128, 8, 32, 16, 16, 1, 4, S<16, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>,
DeviceGemm_Xdl_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 32, 256, 128, 8, 32, 32, 32, 1, 2, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0, 1, 1, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, half_t, half_t, false, true>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_i4_f16/device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_mem_v2_default_instance.cpp 0 → 100644
View file @ 1d8e4ec2
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, I4, F16, PassThrough, PassThrough, PassThrough>>>&
instances)
{
add_device_operation_instances(
instances,
device_gemm_xdl_universal_f16_i4_f16_mk_nk_mn_mem_instances<Interwave, GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profiler/profile_gemm_universal_impl.hpp
View file @ 1d8e4ec2
......@@ -65,11 +65,13 @@ bool profile_gemm_universal_impl(int do_verification,
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<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes();
int rotating_count = std::max(
std::size_t total_gemm_needed =
a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes();
int rotating_count = std::max(
1,
std::min(n_iter,
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
......@@ -86,9 +88,13 @@ bool profile_gemm_universal_impl(int do_verification,
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
break;
default:
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
......@@ -100,11 +106,10 @@ bool profile_gemm_universal_impl(int do_verification,
const auto c_element_op = CElementOp{};
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
using DeviceOp = ck::tensor_operation::device::DeviceGemmV2<ALayout,
BLayout,
......@@ -153,6 +158,84 @@ bool profile_gemm_universal_impl(int do_verification,
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
const int KPerBlock = op_ptr->GetKPerBlock();
if(op_ptr->GetPermuteB())
{
int K1 = KPerBlock;
int K0 = K / KPerBlock;
// int K0, N, K1
for(int j = 0; j < K0; j++)
{
for(int i = 0; i < N; i++)
{
for(int jj = 0; jj < K1; jj++)
{
b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
}
}
}
if(is_same_v<BDataType, pk_i4_t> && is_same_v<ADataType, half_t>)
{
// vector pk_i4x4 permute
for(int i = 0; i < N; i++)
{
for(int j = 0; j < K; j += 8)
{
int input[8];
for(int k = 0; k < 4; k++)
{
int i4x2 = b_k_n_permute(j + k * 2, i);
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
}
// permute 01234567->20643175
{
int hi = input[2];
int lo = input[0];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 0, i) = i4x2;
}
{
int hi = input[6];
int lo = input[4];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 2, i) = i4x2;
}
{
int hi = input[3];
int lo = input[1];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 4, i) = i4x2;
}
{
int hi = input[7];
int lo = input[5];
int i4x2 = (hi << 4) | lo;
b_k_n_permute(j + 6, i) = i4x2;
}
}
}
}
}
else
{
b_k_n_permute = b_k_n;
}
b_device_buf.ToDevice(b_k_n_permute.mData.data());
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
if(KBatch > 0)
......@@ -240,7 +323,15 @@ bool profile_gemm_universal_impl(int do_verification,
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
static constexpr index_t BPackedSize = []() {
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
return 2;
else
return 1;
}();
std::size_t num_btype = sizeof(ADataType) * M * K +
sizeof(BDataType) * K * N / BPackedSize +
sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......
profiler/src/CMakeLists.txt
View file @ 1d8e4ec2
......@@ -177,5 +177,4 @@ if(DL_KERNELS)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif()
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
profiler/src/profile_gemm_universal.cpp
View file @ 1d8e4ec2
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_universal_impl.hpp"
#include "profiler_operation_registry.hpp"
......@@ -27,6 +27,8 @@ enum struct GemmDataType
F16_F8_F16, // 5
F16_F16_F16_F8, // 6
F8_F8_BF16, // 7
F16_I4_F16, // 8
BF16_I4_BF16, // 9
};
#define OP_NAME "gemm_universal"
......@@ -39,7 +41,7 @@ int profile_gemm_universal(int argc, char* argv[])
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
"f16->f8; 7: f8->bf16, "
"comp f8)\n");
"comp f8; 8: f16@i4; 9: bf16@i4\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
......@@ -103,6 +105,7 @@ int profile_gemm_universal(int argc, char* argv[])
using BF16 = ck::bhalf_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t;
using I4 = ck::pk_i4_t;
#endif
using Row = ck::tensor_layout::gemm::RowMajor;
......@@ -207,6 +210,14 @@ int profile_gemm_universal(int argc, char* argv[])
{
return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, I4{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::BF16_I4_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(BF16{}, I4{}, BF16{}, F32{}, BF16{}, Row{}, Col{}, Row{});
}
#endif
else
{
......
script/cmake-ck-dev.sh
View file @ 1d8e4ec2
......@@ -17,7 +17,7 @@ fi
cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \
-D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=ON \
-D GPU_TARGETS=$GPU_TARGETS \
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment