Skip to content

GitLab

Unverified Commit e7e224d3 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-1010

parents fcfe70f9 9fa379ea
Show whitespace changes
Inline Side-by-side
Showing with 768 additions and 79 deletions
CMakeLists.txt
View file @ e7e224d3
......@@ -118,7 +118,7 @@ else()
add_definitions(-DPROFILER_ONLY)
set(GPU_TARGETS "" CACHE STRING "" FORCE)
if(GPU_TARGETS)
message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx90, gfx94, gfx10, or gfx11")
message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx90, gfx94, gfx10, gfx11 or gfx12")
endif()
if(GPU_ARCH MATCHES "gfx90")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx908;gfx90a")
......@@ -128,8 +128,10 @@ else()
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1030")
elseif(GPU_ARCH MATCHES "gfx11")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1100;gfx1101;gfx1102")
elseif(GPU_ARCH MATCHES "gfx12")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1200;gfx1201")
else()
message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx90, gfx94, gfx10, or gfx11")
message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx90, gfx94, gfx10, gfx11 or gfx12")
endif()
set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
endif()
......
example/01_gemm/CMakeLists.txt
View file @ e7e224d3
......@@ -27,7 +27,8 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
if(GPU_TARGETS MATCHES "gfx11")
if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12")
add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
add_example_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
......@@ -74,4 +75,3 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
add_example_executable(example_gemm_xdl_fp16_fp8 gemm_xdl_fp16_fp8.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8)
example/01_gemm/gemm_wmma_fp16.cpp
View file @ e7e224d3
......@@ -19,11 +19,9 @@ using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
< ALayout,
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle<ALayout,
BLayout,
CLayout,
ADataType,
......@@ -35,35 +33,34 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
BElementOp,
CElementOp,
GemmDefault,
1, // Prefetch stage
128, // BlockSize
64, // MPerBlock
128, // NPerBlock
64, // KPerBlock
8, // K1
16, // MPerWmma
16, // NPerWmma
2, // M-Repeat // M-PerWmma / M-Repeat = M-Wave
4, // N-Repeat // N-PerWmma / N-Repeat = N-Wave
S<4, 32, 1>,
1,
32,
16,
32,
64,
8,
16,
16,
1,
2,
S<2, 16, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 32, 1>,
S<2, 16, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1, // C shuffle (M Repeat) Per store
1, // C shuffle (N Repeat) Per store
S<1, 32, 1, 4>,
1,
1,
S<1, 16, 1, 2>,
8>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
......
example/02_gemm_bilinear/CMakeLists.txt
View file @ e7e224d3
list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102 gfx1103)
list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
......
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View file @ e7e224d3
add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx11")
if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12")
add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
endif()
example/64_fpAintB_gemm/CMakeLists.txt
View file @ e7e224d3
if(GPU_TARGETS MATCHES "gfx11")
if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12")
add_custom_target(example_fpAintB_gemm_wmma)
add_example_executable(example_fp16int8_gemm_wmma fp16int8_gemm_wmma.cpp)
add_dependencies(example_fpAintB_gemm_wmma example_fp16int8_gemm_wmma)
......
include/ck/ck.hpp
View file @ e7e224d3
......@@ -104,13 +104,6 @@
#define CK_USE_AMD_MFMA_GFX940
#endif
// WMMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_WMMA
#elif defined(__gfx11__) // || defined(__gfx12__) // for GPU code
#define CK_USE_AMD_WMMA
#endif
// buffer load
#define CK_USE_AMD_BUFFER_LOAD 1
......
include/ck/host_utility/device_prop.hpp
View file @ e7e224d3
......@@ -84,8 +84,10 @@ inline bool is_navi3_supported()
return ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102" || ck::get_device_name() == "gfx1103";
}
inline bool is_navi4_supported()
{
return ck::get_device_name() == "gfx1200" || ck::get_device_name() == "gfx1201";
}
} // namespace ck
include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
View file @ e7e224d3
......@@ -13,6 +13,475 @@
namespace ck {
#ifdef __gfx12__
template <index_t BlockSize,
typename FloatA,
typename FloatB,
typename FloatAcc,
typename ABlockDesc,
typename BBlockDesc,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerWMMA,
index_t NPerWMMA,
index_t MRepeat,
index_t NRepeat,
index_t KPack,
bool AEnableLds = true,
bool BEnableLds = true,
bool TransposeC = false>
/* Option: Read from LDS, big buffer hold all threads required data
* Source
* A: K0PerBlock x MPerBlock x K1
* B: K0PerBlock x NPerBlock x K1
* Destination
* C, non-transpose
* thread level: MRepeat x NRepeat x MAccVgprs
* block level: MRepeat x MWave x MSubGroup x NRepeat x NWave x NThreadPerSubGroup x MAccVgprs
* KPACK == WMMA_K = 16
*
* Option: Read from VMEM, small buffer hold each thread own required data (Skip LDS)
* Source:
* A(if skip LDS): MRepeat x KPack
* B(if skip LDS): NRepeat x KPack
* Destination
* C, non-transpose
* block level: MRepeat x MWave x MSubGroup x NRepeat x NWave x NThreadPerSubGroup x MAccVgprs
*/
struct BlockwiseGemmWMMA
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto WmmaK = Number<16>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
// Hardcode of WaveSize, since current HIP Runtime(5.4.0-10984) could not return correct one.
static constexpr index_t WaveSize = 32;
// When use LDS, each Row(16 consecutive lanes) read whole data from source buffer
// When not use LDS, each Row read half of whole data from source buffer, exchange the data via
// permutation
static constexpr index_t A_KRow = 2;
static constexpr index_t B_KRow = 2;
static constexpr index_t A_K1 = ABlockDesc{}.GetLength(I5);
static constexpr index_t B_K1 = BBlockDesc{}.GetLength(I5);
static constexpr auto wmma_gemm =
WmmaGemm<FloatA, FloatB, FloatAcc, MPerWMMA, NPerWMMA, KPack, TransposeC>{};
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerWMMA);
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerWMMA);
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
FloatAcc,
MRepeat * NRepeat,
wmma_gemm.GetRegSizePerWmma(),
true>
c_thread_buf_;
__host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
__device__ static auto GetWaveIdx()
{
const index_t thread_id = ThisThreadBlock::GetThreadId();
constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
// Default, Block buffer in LDS, thread level offset enabled
__device__ static auto CalculateAThreadOriginDataIndex()
{
if constexpr(AEnableLds)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto WMMA_a_idx = wmma_gemm.CalculateAThreadOriginDataIndex();
// |KRepeat |MRepeat|MWave |KRow |MLane |KPack
return make_tuple(0, 0, waveId_m, wmma_gemm.GetSubGroupId(), WMMA_a_idx, 0);
}
else
{
return make_tuple(0, 0, 0, 0, 0, 0);
}
}
__device__ static auto CalculateBThreadOriginDataIndex()
{
if constexpr(BEnableLds)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_n = wave_idx[I1];
const auto WMMA_b_idx = wmma_gemm.CalculateBThreadOriginDataIndex();
// |KRepeat |NRepeat|Nwave |KRow |NLane |KPack
return make_tuple(0, 0, waveId_n, wmma_gemm.GetSubGroupId(), WMMA_b_idx, 0);
}
else
{
return make_tuple(0, 0, 0, 0, 0, 0);
}
}
template <index_t m0, index_t n0>
__device__ static auto CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const auto blk_idx = wmma_gemm.GetBeginOfThreadBlk();
constexpr auto mrepeat_mwave_mperWMMA_to_m_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerWMMA))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1, 2>{}));
constexpr auto nrepeat_nwave_nperWMMA_to_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerWMMA))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1, 2>{}));
const index_t c_thread_m = mrepeat_mwave_mperWMMA_to_m_adaptor.CalculateBottomIndex(
make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
const index_t c_thread_n = nrepeat_nwave_nperWMMA_to_n_adaptor.CalculateBottomIndex(
make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
return make_tuple(c_thread_m, c_thread_n);
}
template <index_t m0, index_t n0>
__device__ static auto CalculateCThreadOriginDataIndex7D(Number<m0>, Number<n0>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const auto blk_idx = wmma_gemm.GetBeginOfThreadBlk3D();
return make_tuple(
Number<m0>{}, waveId_m, blk_idx[I0], Number<n0>{}, waveId_n, blk_idx[I1], blk_idx[I2]);
}
using Tuple6 = decltype(CalculateAThreadOriginDataIndex());
__host__ __device__ BlockwiseGemmWMMA(Tuple6 a_origin = CalculateAThreadOriginDataIndex(),
Tuple6 b_origin = CalculateBThreadOriginDataIndex())
: a_thread_copy_(a_origin), b_thread_copy_(b_origin)
{
static_assert(ABlockDesc::IsKnownAtCompileTime() && BBlockDesc::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
"ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
static_assert(MPerBlock % (MPerWMMA * MRepeat) == 0 &&
NPerBlock % (NPerWMMA * NRepeat) == 0,
"wrong!");
static_assert(AEnableLds == true, "only support EnableLds");
static_assert(BEnableLds == true, "only support EnableLds");
}
// transposed WMMA output C' = B' * A'
__host__ __device__ static constexpr auto
GetCThreadDescriptor_MRepeat_MWave_MThreadPerSubGroup_NRepeat_NWave_NSubGroup_NAccVgprs()
{
constexpr auto c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens =
wmma_gemm.GetCMSubGroupNThreadPerSubGroupMAccVgprsThreadBlkLengths();
constexpr auto NAccVgprs = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I2];
return make_naive_tensor_descriptor_packed(
// |MRepeat |MWave |MSubGroup |NRepeat |NWave
// |NThreadPerSubGroup |MAccVgprs
make_tuple(Number<MRepeat>{}, I1, I1, Number<NRepeat>{}, I1, I1, NAccVgprs));
}
// Thread level, register decriptor. Vector-write
__host__ __device__ static constexpr auto
GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs()
{
constexpr auto c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens =
wmma_gemm.GetCMSubGroupNThreadPerSubGroupMAccVgprsThreadBlkLengths();
constexpr auto MAccVgprs = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I2];
constexpr auto AccStride = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I3];
return make_naive_tensor_descriptor(
// |MRepeat |MWave |MSubGroup |NRepeat |NWave
// |NThreadPerSubGroup |MAccVgprs
make_tuple(Number<MRepeat>{}, I1, I1, Number<NRepeat>{}, I1, I1, MAccVgprs),
make_tuple(Number<NRepeat>{} * MAccVgprs * AccStride,
Number<NRepeat>{} * MAccVgprs * AccStride,
Number<NRepeat>{} * MAccVgprs * AccStride,
MAccVgprs * AccStride,
MAccVgprs * AccStride,
MAccVgprs * AccStride,
AccStride));
}
template <typename CGridDesc_M_N>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
const CGridDesc_M_N& c_grid_desc_m_n)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
const auto c_grid_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma =
transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(
make_unmerge_transform(make_tuple(M / (MWaves * MPerWMMA), MWaves, MPerWMMA)),
make_unmerge_transform(make_tuple(N / (NWaves * NPerWMMA), NWaves, NPerWMMA))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
return wmma_gemm
.MakeCDesc_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
c_grid_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma);
}
// transposed WMMA output C' = B' * A'
__host__ __device__ static constexpr auto
GetCBlockDescriptor_MRepeat_MWave_MThreadPerSubGroup_NRepeat_NWave_NSubGroup_NAccVgprs()
{
constexpr auto c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<MWaves>{},
Number<MPerWMMA>{},
Number<NRepeat>{},
Number<NWaves>{},
Number<NPerWMMA>{}));
return wmma_gemm
.MakeCDesc_MBlockxRepeat_MWave_MThreadPerSubGroup_NBlockxRepeat_NWave_NSubGroup_NAccVgprs(
c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma);
}
// Provide dimension size
__host__ __device__ static constexpr auto
GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs()
{
constexpr auto c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<MWaves>{},
Number<MPerWMMA>{},
Number<NRepeat>{},
Number<NWaves>{},
Number<NPerWMMA>{}));
return wmma_gemm
.MakeCDesc_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma);
}
// Describe how data allocated in thread copy src buffer
// M0_M1_M2 = MRepeat_MWave_MPerWmma, N0_N1_N2 = NRepeat_NWave_NPerWmma
static constexpr ABlockDesc a_block_desc_k0_m0_m1_m2_k1;
static constexpr BBlockDesc b_block_desc_k0_n0_n1_n2_k1;
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
__device__ void Run(const ABlockBuffer& a_block_buf,
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
b_thread_desc_.GetElementSpaceSize());
static_assert(KPack % (A_K1 * A_KRow) == 0, "");
static_assert(KPack % (B_K1 * B_KRow) == 0, "");
// basic intrinsic to determine loopover direction
if constexpr(MRepeat < NRepeat)
{
static_for<0, KPerBlock / KPack, 1>{}(
[&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A
a_thread_copy_.Run(
a_block_desc_k0_m0_m1_m2_k1,
make_tuple(Number<k * KPack / A_K1>{}, m0, I0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, m0, I0, I0, I0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B
b_thread_copy_.Run(
b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<k * KPack / B_K1>{}, n0, I0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, n0, I0, I0, I0, I0),
b_thread_buf);
vector_type<FloatA, KPack / A_KRow> a_thread_vec;
vector_type<FloatB, KPack / B_KRow> b_thread_vec;
static_for<0, KPack / A_KRow, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatA>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(i / A_K1, m0, 0, 0, 0, i % A_K1))>{}];
});
static_for<0, KPack / B_KRow, 1>{}([&](auto i) {
b_thread_vec.template AsType<FloatB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(i / B_K1, n0, 0, 0, 0, i % B_K1))>{}];
});
using wmma_input_type_a =
typename vector_type<FloatA, WmmaK / A_KRow>::type;
using wmma_input_type_b =
typename vector_type<FloatB, WmmaK / B_KRow>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type_a>(),
b_thread_vec.template AsType<wmma_input_type_b>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
else
{
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KPerBlock / KPack, 1>{}([&](auto k) { // k=0,1,2 instead of
// k=0,kpack*1, ..
// read B
b_thread_copy_.Run(
b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<k * KPack / B_K1>{}, n0, I0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, n0, I0, I0, I0, I0),
b_thread_buf);
// read A
a_thread_copy_.Run(
a_block_desc_k0_m0_m1_m2_k1,
make_tuple(Number<k * KPack / A_K1>{}, m0, I0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, m0, I0, I0, I0, I0),
a_thread_buf);
vector_type<FloatA, KPack / A_KRow> a_thread_vec;
vector_type<FloatB, KPack / B_KRow> b_thread_vec;
static_for<0, KPack / A_KRow, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatA>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(i / A_K1, m0, 0, 0, 0, i % A_K1))>{}];
});
static_for<0, KPack / B_KRow, 1>{}([&](auto i) {
b_thread_vec.template AsType<FloatB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(i / B_K1, n0, 0, 0, 0, i % B_K1))>{}];
});
using wmma_input_type_a =
typename vector_type<FloatA, WmmaK / A_KRow>::type;
using wmma_input_type_b =
typename vector_type<FloatB, WmmaK / B_KRow>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type_a>(),
b_thread_vec.template AsType<wmma_input_type_b>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
}
protected:
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<KPack / A_K1 / A_KRow>{}, Number<MRepeat>{}, I1, I1, I1, Number<A_K1>{}),
make_tuple(Number<A_K1>{},
Number<KPack / A_KRow>{},
Number<A_K1>{},
Number<A_K1>{},
Number<A_K1>{},
Number<1>{}));
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<KPack / B_K1 / B_KRow>{}, Number<NRepeat>{}, I1, I1, I1, Number<B_K1>{}),
make_tuple(Number<B_K1>{},
Number<KPack / B_KRow>{},
Number<B_K1>{},
Number<B_K1>{},
Number<B_K1>{},
Number<1>{}));
// C[M, N, NumRegWMMA]
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, wmma_gemm.GetRegSizePerWmma()));
template <bool EnableLds>
struct AThreadCopySelector;
template <>
struct AThreadCopySelector<true>
{
using type =
ThreadwiseTensorSliceTransfer_v4<FloatA,
FloatA,
decltype(a_block_desc_k0_m0_m1_m2_k1),
decltype(a_thread_desc_),
Sequence<KPack / A_K1 / A_KRow, 1, 1, 1, 1, A_K1>,
Sequence<0, 1, 2, 3, 4, 5>,
5,
A_K1,
A_K1>;
};
template <bool EnableLds>
struct BThreadCopySelector;
template <>
struct BThreadCopySelector<true>
{
using type =
ThreadwiseTensorSliceTransfer_v4<FloatB,
FloatB,
decltype(b_block_desc_k0_n0_n1_n2_k1),
decltype(b_thread_desc_),
Sequence<KPack / B_K1 / B_KRow, 1, 1, 1, 1, B_K1>,
Sequence<0, 1, 2, 3, 4, 5>,
5,
B_K1,
B_K1>;
};
typename AThreadCopySelector<AEnableLds>::type a_thread_copy_;
typename BThreadCopySelector<BEnableLds>::type b_thread_copy_;
};
#else
template <index_t BlockSize,
typename FloatA,
typename FloatB,
......@@ -68,6 +537,7 @@ struct BlockwiseGemmWMMA
// permutation
static constexpr index_t A_KRow = AEnableLds ? 1 : 2;
static constexpr index_t B_KRow = BEnableLds ? 1 : 2;
static constexpr index_t A_K1 = ABlockDesc{}.GetLength(I5);
static constexpr index_t B_K1 = BBlockDesc{}.GetLength(I5);
......@@ -529,5 +999,6 @@ struct BlockwiseGemmWMMA
typename AThreadCopySelector<AEnableLds>::type a_thread_copy_;
typename BThreadCopySelector<BEnableLds>::type b_thread_copy_;
};
#endif
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
View file @ e7e224d3
......@@ -137,8 +137,8 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
static constexpr auto BEnableLds_auto = MWaves == 1 ? false : true;
// If true, LDS is used unconditionally
static constexpr auto AEnableLds_manu = false;
static constexpr auto BEnableLds_manu = false;
static constexpr auto AEnableLds_manu = true;
static constexpr auto BEnableLds_manu = true;
static constexpr auto AEnableLds = AEnableLds_auto || AEnableLds_manu || (NumPrefetch > 1);
static constexpr auto BEnableLds = BEnableLds_auto || BEnableLds_manu || (NumPrefetch > 1);
......
include/ck/tensor_operation/gpu/device/impl/device_fpAintB_gemm_wmma.hpp
View file @ e7e224d3
......@@ -509,7 +509,7 @@ struct DeviceFpAintBGemm_Wmma_CShuffle : public DeviceGemm_dequantB<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::is_navi3_supported())
if(ck::is_navi3_supported() || ck::is_navi4_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, ck::half_t> ||
is_same_v<AccDataType, int32_t>))
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
View file @ e7e224d3
......@@ -101,8 +101,8 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
(MWaves == 1 && is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value) ? false : true;
// If true, LDS is used unconditionally
static constexpr auto AEnableLds_manu = false;
static constexpr auto BEnableLds_manu = false;
static constexpr auto AEnableLds_manu = true;
static constexpr auto BEnableLds_manu = true;
static constexpr auto AEnableLds = AEnableLds_auto || AEnableLds_manu || (NumPrefetch > 1);
static constexpr auto BEnableLds = BEnableLds_auto || BEnableLds_manu || (NumPrefetch > 1);
......@@ -515,7 +515,7 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::is_navi3_supported())
if(ck::is_navi3_supported() || ck::is_navi4_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
View file @ e7e224d3
......@@ -94,8 +94,8 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
(MWaves == 1 && is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value) ? false : true;
// If true, LDS is used unconditionally
static constexpr auto AEnableLds_manu = false;
static constexpr auto BEnableLds_manu = false;
static constexpr auto AEnableLds_manu = true;
static constexpr auto BEnableLds_manu = true;
static constexpr auto AEnableLds = AEnableLds_auto || AEnableLds_manu || (NumPrefetch > 1);
static constexpr auto BEnableLds = BEnableLds_auto || BEnableLds_manu || (NumPrefetch > 1);
......@@ -443,7 +443,7 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::is_navi3_supported())
if(ck::is_navi3_supported() || ck::is_navi4_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, ck::half_t> ||
is_same_v<AccDataType, int32_t>))
......
include/ck/tensor_operation/gpu/grid/gridwise_fpAintB_gemm_wmma.hpp
View file @ e7e224d3
......@@ -50,8 +50,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
defined(__gfx1102__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
__shared__ char p_shared[GridwiseGemm::SharedMemTrait::lds_size];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
View file @ e7e224d3
......@@ -54,7 +54,7 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -147,7 +147,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2CTileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
// printf("entry kernel launch");
__shared__ char p_shared[GridwiseOp::SharedMemTrait::lds_size];
......@@ -237,7 +237,7 @@ __global__ void
const CDEElementwiseOperation cde_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
__shared__ char p_shared[GridwiseOp::SharedMemTrait::lds_size];
GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
View file @ e7e224d3
......@@ -45,7 +45,7 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
__shared__ char p_shared[GridwiseGemm::SharedMemTrait::lds_size];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
View file @ e7e224d3
......@@ -11,12 +11,17 @@ namespace ck {
enum struct WmmaInstr
{
// gfx11
wmma_f32_16x16x16_f16 = 0,
wmma_f32_16x16x16_bf16,
wmma_f16_16x16x16_f16,
wmma_bf16_16x16x16_bf16,
wmma_i32_16x16x16_iu8,
wmma_i32_16x16x16_iu4
wmma_i32_16x16x16_iu4,
// gfx12
wmma_f32_16x16x16_f16_gfx12,
wmma_f32_16x16x16_bf16_gfx12,
wmma_i32_16x16x16_iu8_gfx12,
};
/*
......@@ -279,6 +284,122 @@ struct wmma_type<WmmaInstr::wmma_i32_16x16x16_iu8,
}
};
// gfx12
// A-swizzled
template <index_t WaveSize>
struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16_gfx12,
WaveSize,
typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
{
// Absolute fixing property
// * Data Pixel
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
// static constexpr index_t src_a_data_size = 2;
// static constexpr index_t src_b_data_size = 2;
// static constexpr index_t acc_data_size = 4;
// * Thread mapping inside wave, num_thread_per_subgroups always alone N direction
static constexpr index_t acc_data_size = 4;
static constexpr index_t acc_pack_number = 1;
static constexpr index_t num_thread_per_subgroups = n_per_wmma;
// Wave mode dependent propety
static constexpr index_t wave_size = Number<WaveSize>{};
// * Fixed in Navi3x, Will be wave mode dependent on Navi4x
// static constexpr index_t num_src_a_vgprs_per_wave = k_per_wmma / 2 * src_a_data_size / 4;
// static constexpr index_t num_src_b_vgprs_per_wave = k_per_wmma / 2 * src_b_data_size / 4;
// * num_acc_vgprs_per_wave alone M direction
// * num_subgroups alone M direction
static constexpr index_t num_acc_vgprs_per_wave = m_per_wmma * n_per_wmma / wave_size;
static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
template <index_t MPerWmma, index_t NPerWmma, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
static_assert(wave_size == 32, "only support wave32 for gfx12 wmma");
if constexpr(wave_size == 32)
{
intrin_wmma_f32_16x16x16_f16_w32_gfx12<MPerWmma, NPerWmma>::Run(a, b, reg_c);
}
}
};
template <index_t WaveSize>
struct wmma_type<WmmaInstr::wmma_f32_16x16x16_bf16_gfx12,
WaveSize,
typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
{
// Absolute fixing property
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
// static constexpr index_t src_a_data_size = 2;
// static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 4;
static constexpr index_t acc_pack_number = 1;
static constexpr index_t num_thread_per_subgroups = n_per_wmma;
// Wave mode dependent propety
static constexpr index_t wave_size = Number<WaveSize>{};
// static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
// static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
static constexpr index_t num_acc_vgprs_per_wave = m_per_wmma * n_per_wmma / wave_size;
static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
template <index_t MPerWmma, index_t NPerWmma, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
static_assert(wave_size == 32, "only support wave32 for gfx12 wmma");
if constexpr(wave_size == 32)
{
intrin_wmma_f32_16x16x16_bf16_w32_gfx12<MPerWmma, NPerWmma>::Run(a, b, reg_c);
}
}
};
template <index_t WaveSize>
struct wmma_type<WmmaInstr::wmma_i32_16x16x16_iu8_gfx12,
WaveSize,
typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
{
// Absolute fixing property
static constexpr index_t m_per_wmma = 16;
static constexpr index_t n_per_wmma = 16;
static constexpr index_t k_per_wmma = 16;
// static constexpr index_t src_a_data_size = 2;
// static constexpr index_t src_b_data_size = 2;
static constexpr index_t acc_data_size = 4;
static constexpr index_t acc_pack_number = 1;
static constexpr index_t num_thread_per_subgroups = n_per_wmma;
// Wave mode dependent propety
static constexpr index_t wave_size = Number<WaveSize>{};
// static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
// static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
static constexpr index_t num_acc_vgprs_per_wave = m_per_wmma * n_per_wmma / wave_size;
static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
template <index_t MPerWmma,
index_t NPerWmma,
class FloatA,
class FloatB,
class FloatC,
bool neg_a = false,
bool neg_b = false,
bool clamp = false>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
static_assert(wave_size == 32, "only support wave32 for gfx12 wmma");
if constexpr(wave_size == 32)
{
intrin_wmma_i32_16x16x16_iu8_w32_gfx12<MPerWmma, NPerWmma, neg_a, neg_b, clamp>::Run(
a, b, reg_c);
}
}
};
template <typename src_type_a,
typename src_type_b,
typename dst_type,
......@@ -296,13 +417,21 @@ struct WmmaSelector
template <>
static constexpr auto GetWmma<half_t, half_t, float, 16, 16>()
{
#ifdef __gfx12__
return WmmaInstr::wmma_f32_16x16x16_f16_gfx12;
#else
return WmmaInstr::wmma_f32_16x16x16_f16;
#endif
}
template <>
static constexpr auto GetWmma<bhalf_t, bhalf_t, float, 16, 16>()
{
#ifdef __gfx12__
return WmmaInstr::wmma_f32_16x16x16_bf16_gfx12;
#else
return WmmaInstr::wmma_f32_16x16x16_bf16;
#endif
}
template <>
......@@ -320,8 +449,13 @@ struct WmmaSelector
template <>
static constexpr auto GetWmma<int8_t, int8_t, int, 16, 16>()
{
#ifdef __gfx12__
return WmmaInstr::wmma_i32_16x16x16_iu8_gfx12;
#else
return WmmaInstr::wmma_i32_16x16x16_iu8;
#endif
}
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
template <>
static constexpr auto GetWmma<int4_t, int4_t, int, 16, 16>()
......@@ -502,6 +636,9 @@ struct WmmaGemm
__device__ static auto GetSubGroupId()
{
static_assert(wmma_instr.num_thread_per_subgroups * wmma_instr.num_subgroups ==
wmma_instr.wave_size,
"");
return (GetLaneId() / wmma_instr.num_thread_per_subgroups) % wmma_instr.num_subgroups;
}
......@@ -516,12 +653,20 @@ struct WmmaGemm
__host__ __device__ static auto CalculateAThreadOriginDataIndex()
{
#ifdef __gfx12__
return GetLaneIdUnderSubGroup();
#else
return TransposeC ? GetLaneIdUnderSubGroup() : GetSwizzledLaneIdLow();
#endif
}
__host__ __device__ static auto CalculateBThreadOriginDataIndex()
{
#ifdef __gfx12__
return GetLaneIdUnderSubGroup();
#else
return TransposeC ? GetSwizzledLaneIdLow() : GetLaneIdUnderSubGroup();
#endif
}
__device__ static CIndex GetBeginOfThreadBlk()
......
include/ck/utility/amd_wmma.hpp
View file @ e7e224d3
......@@ -257,5 +257,87 @@ struct intrin_wmma_i32_16x16x16_iu8_w64<16, 16, neg_a, neg_b, clamp>
}
};
// gfx12
/********************************WAVE32 MODE***********************************************/
#if defined(__gfx1200__)
#define __gfx12__
#endif
// src: fp16, dst: fp32
template <index_t MPerWave, index_t NPerWave>
struct intrin_wmma_f32_16x16x16_f16_w32_gfx12;
template <>
struct intrin_wmma_f32_16x16x16_f16_w32_gfx12<16, 16>
{
template <class FloatC>
__device__ static void Run(const half8_t& reg_a, const half8_t& reg_b, FloatC& reg_c)
{
// * Inline assembly need to elimate the duplicated data load, compiler won't help you
// delete them.
// amd_assembly_wmma_f32_16x16x16_f16_w32(
// reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
#if defined(__gfx12__)
reg_c.template AsType<float8_t>()(Number<0>{}) =
__builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(
reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif
}
};
// src: bf16, dst: fp32
template <index_t MPerWave, index_t NPerWave>
struct intrin_wmma_f32_16x16x16_bf16_w32_gfx12;
template <>
struct intrin_wmma_f32_16x16x16_bf16_w32_gfx12<16, 16>
{
template <class FloatC>
__device__ static void Run(const bhalf8_t& reg_a, const bhalf8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx12__)
reg_c.template AsType<float8_t>()(Number<0>{}) =
__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(
reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif
}
};
// src: iu8, dst: i32
template <index_t MPerWave, index_t NPerWave, bool neg_a, bool neg_b, bool clamp>
struct intrin_wmma_i32_16x16x16_iu8_w32_gfx12;
template <bool neg_a, bool neg_b, bool clamp>
struct intrin_wmma_i32_16x16x16_iu8_w32_gfx12<16, 16, neg_a, neg_b, clamp>
{
template <class FloatC>
__device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
{
#if defined(__gfx12__)
reg_c.template AsType<int32x8_t>()(Number<0>{}) =
__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
neg_a,
bit_cast<int32x2_t>(reg_a),
neg_b,
bit_cast<int32x2_t>(reg_b),
reg_c.template AsType<int32x8_t>()[Number<0>{}],
clamp);
#else
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif
}
};
} // namespace ck
#endif
profiler/src/CMakeLists.txt
View file @ e7e224d3
......@@ -108,8 +108,6 @@ if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
endif()
if(DL_KERNELS)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
endif()
......
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