Skip to content

GitLab

Unverified Commit de37550f authored by Anthony Chang's avatar Anthony Chang Committed by GitHub
Browse files

Input/output permutation for fused attention (#460) 



* reopen masking att instance due to CI is upgraded

* re-enable instances previously failed on 9110

* enable ksize-kpadding pair validity test

* add non-masked attention+permute test; expose masking boolean to attention kernel handles

* disable bench

* fix test

* move files

* bulk rename batched_gemm_masking_scale_softmax_gemm_permute to batched_gemm_softmax_gemm_permute

* format

* amend rename

* disable bench in test

* add mask/no-mask test for non-permute attention kernels

* disable broken kernel instance

* example working

add non-permuted problem statement

evaluating whether overhead comes from permutation or the extra kernel arg

* interface for bias addition without implementing it

* test and profiler running

* tidy

* mask type determined by enum class

* unify example code

* move masking specialization to its own header

* align formats

* extract helper functions

* experiment merging dims for attn w/ permute; shows perf parity with attn wo/ permute

* add tensor specialization to template args

since tensor spec packed shows perf parity when permutation isn't needed

remove redundant template args

comment on 'packed' tensor specialization

* grouped attention with input/output permute example

* format

* clean up

* refactor acc0 tile visitor
Co-authored-by: wangshaojie6's avatarshaojiewang <wsjmessi@163.com>
Co-authored-by: default avatarChao Liu <chao.liu2@amd.com>
parent cd517326
Show whitespace changes
Inline Side-by-side
Showing with 930 additions and 365 deletions
include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp 0 → 100644
View file @ de37550f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
namespace ck {
namespace tensor_operation {
// assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
device::TensorSpecialization TensorSpec>
static auto MakeGridDescriptorPair(const std::vector<index_t>& gs_ms_ns_lengths_vec,
const std::vector<index_t>& gs_ms_ns_strides_vec)
{
if(!(gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN))
{
throw std::runtime_error("wrong! dimension must match input lengths");
}
const auto to_tuple = [&](auto& vec, auto start, auto end) {
return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
};
const auto gs_ms_ns_lengths =
to_tuple(gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
const auto gs_ms_ns_strides =
to_tuple(gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
// dimension Ids for G0, G1, ...
constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
// dimension Ids for M0, M1, ...
constexpr auto mDimIds =
typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
// dimension Ids for N0, N1, ...
constexpr auto nDimIds =
typename arithmetic_sequence_gen<NumDimG + NumDimM, NumDimG + NumDimM + NumDimN, 1>::type{};
// lengths for G0, G1, ...
const auto gLengths = get_container_subset(gs_ms_ns_lengths, gDimIds);
// lengths for M0, M1, ...
const auto mLengths = get_container_subset(gs_ms_ns_lengths, mDimIds);
// lengths for N0, N1, ...
const auto nLengths = get_container_subset(gs_ms_ns_lengths, nDimIds);
if constexpr(TensorSpec == device::TensorSpecialization::Packed)
{
auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
const auto grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
make_tuple(G, M, N),
make_tuple(gs_ms_ns_strides[Number<NumDimG - 1>{}],
gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
const auto grid_desc_mraw_nraw = make_naive_tensor_descriptor(
make_tuple(M, N),
make_tuple(gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
return std::make_pair(grid_desc_g_mraw_nraw, grid_desc_mraw_nraw);
}
else
{
// naive tensor C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
const auto grid_desc_gs_ms_ns =
make_naive_tensor_descriptor(gs_ms_ns_lengths, gs_ms_ns_strides);
// transformed tensor C[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
// N2 * ...]
// Note: This does not require padding as it only provides G offset calculation. Technically
// descriptor for only G is needed. Here we opt for backward compatibility purpose to return
// G_M_N
const auto grid_desc_g_mraw_nraw =
transform_tensor_descriptor(grid_desc_gs_ms_ns,
make_tuple(make_merge_transform(gLengths),
make_merge_transform(mLengths),
make_merge_transform(nLengths)),
make_tuple(gDimIds, mDimIds, nDimIds),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
const auto c_ms_ns_lengths = to_tuple(
gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
const auto c_ms_ns_strides = to_tuple(
gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
// transformed tensor C[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
// N2 * ...]
const auto grid_desc_ms_ns = make_naive_tensor_descriptor(c_ms_ns_lengths, c_ms_ns_strides);
const auto grid_desc_mraw_nraw = transform_tensor_descriptor(
grid_desc_ms_ns,
make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
make_tuple(mDimIds - Number<NumDimG>{}, nDimIds - Number<NumDimG>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return std::make_pair(grid_desc_g_mraw_nraw, grid_desc_mraw_nraw);
}
}
template <typename NumDims_G_M_N_K_O, // Sequence<>
typename PerBlock_M_N_K_O, // Sequence<>
device::GemmSpecialization GemmSpec,
device::TensorSpecialization ASpec,
device::TensorSpecialization B0Spec,
device::TensorSpecialization B1Spec,
device::TensorSpecialization CSpec>
struct TransformBatchedContractionContractionToBatchedGemmGemm
{
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 index_t NumDimG = NumDims_G_M_N_K_O::At(I0);
static constexpr index_t NumDimM = NumDims_G_M_N_K_O::At(I1);
static constexpr index_t NumDimN = NumDims_G_M_N_K_O::At(I2);
static constexpr index_t NumDimK = NumDims_G_M_N_K_O::At(I3);
static constexpr index_t NumDimO = NumDims_G_M_N_K_O::At(I4);
static constexpr index_t MPerBlock = PerBlock_M_N_K_O::At(I0);
static constexpr index_t NPerBlock = PerBlock_M_N_K_O::At(I1);
static constexpr index_t KPerBlock = PerBlock_M_N_K_O::At(I2);
static constexpr index_t OPerBlock = PerBlock_M_N_K_O::At(I3);
static constexpr auto matrix_padder =
device::GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
MPerBlock, NPerBlock, KPerBlock, OPerBlock};
//
// A
//
static auto MakeAGridDescriptorPair(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
const std::vector<index_t>& a_gs_ms_ks_strides_vec)
{
return MakeGridDescriptorPair<NumDimG, NumDimM, NumDimK, ASpec>(a_gs_ms_ks_lengths_vec,
a_gs_ms_ks_strides_vec);
}
// TODO: rename to G_MRaw_KRaw
static auto MakeAGridDescriptor_G_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
const std::vector<index_t>& a_gs_ms_ks_strides_vec)
{
return MakeAGridDescriptorPair(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec).first;
}
static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
const std::vector<index_t>& a_gs_ms_ks_strides_vec)
{
return matrix_padder.PadADescriptor_M_K(
MakeAGridDescriptorPair(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec).second);
}
template <typename AGridDesc_M_K, typename Number>
__host__ __device__ static constexpr auto
MakeAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k, const Number& AK1)
{
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
const auto AK0 = K / AK1;
return transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
//
// B (alias of B0)
//
static auto MakeB0GridDescriptorPair(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
{
return MakeGridDescriptorPair<NumDimG, NumDimN, NumDimK, B0Spec>(b0_gs_ns_ks_lengths_vec,
b0_gs_ns_ks_strides_vec);
}
// TODO: rename to G_MRaw_NRaw
static auto MakeB0GridDescriptor_G_N_K(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
{
return MakeB0GridDescriptorPair(b0_gs_ns_ks_lengths_vec, b0_gs_ns_ks_strides_vec).first;
}
static auto MakeB0GridDescriptor_N_K(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
{
// alias of matrix_padder.PadB0Descriptor_N_K
return matrix_padder.PadBDescriptor_N_K(
MakeB0GridDescriptorPair(b0_gs_ns_ks_lengths_vec, b0_gs_ns_ks_strides_vec).second);
}
template <typename BGridDesc_N_K, typename Number>
__host__ __device__ static constexpr auto
MakeB0GridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k, const Number& BK1)
{
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
const auto BK0 = K / BK1;
return transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
//
// B1
//
static auto MakeB1GridDescriptorPair(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
const std::vector<index_t>& b1_gs_os_ns_strides_vec)
{
return MakeGridDescriptorPair<NumDimG, NumDimO, NumDimN, B1Spec>(b1_gs_os_ns_lengths_vec,
b1_gs_os_ns_strides_vec);
}
// TODO: rename to G_NRaw_KRaw
static auto MakeB1GridDescriptor_G_N_K(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
const std::vector<index_t>& b1_gs_os_ns_strides_vec)
{
return MakeB1GridDescriptorPair(b1_gs_os_ns_lengths_vec, b1_gs_os_ns_strides_vec).first;
}
static auto MakeB1GridDescriptor_N_K(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
const std::vector<index_t>& b1_gs_os_ns_strides_vec)
{
// alias of matrix_padder.PadB1Descriptor_O_N
return matrix_padder.PadB1Descriptor_N_K(
MakeB1GridDescriptorPair(b1_gs_os_ns_lengths_vec, b1_gs_os_ns_strides_vec).second);
}
template <typename B1GridDesc_N_K, typename Number>
__host__ __device__ static constexpr auto
MakeB1GridDescriptor_BK0_N_BK1(const B1GridDesc_N_K& b1_grid_desc_n_k, const Number& B1K1)
{
const auto N = b1_grid_desc_n_k.GetLength(I0);
const auto K = b1_grid_desc_n_k.GetLength(I1);
const auto B1K0 = K / B1K1;
return transform_tensor_descriptor(
b1_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
//
// C
//
static auto MakeCGridDescriptorPair(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
const std::vector<index_t>& c_gs_ms_os_strides_vec)
{
return MakeGridDescriptorPair<NumDimG, NumDimM, NumDimO, CSpec>(c_gs_ms_os_lengths_vec,
c_gs_ms_os_strides_vec);
}
// TODO: rename to G_MRaw_NRaw
static auto MakeCGridDescriptor_G_M_N(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
const std::vector<index_t>& c_gs_ms_os_strides_vec)
{
return MakeCGridDescriptorPair(c_gs_ms_os_lengths_vec, c_gs_ms_os_strides_vec).first;
}
static auto MakeCGridDescriptor_M_N(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
const std::vector<index_t>& c_gs_ms_os_strides_vec)
{
return matrix_padder.PadCDescriptor_M_N(
MakeCGridDescriptorPair(c_gs_ms_os_lengths_vec, c_gs_ms_os_strides_vec).second);
}
};
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
View file @ de37550f
......@@ -28,9 +28,26 @@ void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_g
F16,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
PassThrough>>>& instances);
false>>>& instances);
void add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
Col,
Row,
Row,
F16,
F16,
F16,
F16,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
true>>>& instances);
template <typename ALayout,
typename B0Layout,
......@@ -39,7 +56,8 @@ template <typename ALayout,
typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType>
typename CDataType,
bool MaskOutUpperTriangle>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm<ALayout,
B0Layout,
......@@ -51,9 +69,10 @@ struct DeviceOperationInstanceFactory<
CDataType,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
PassThrough>>
MaskOutUpperTriangle>>
{
using DeviceOp = DeviceBatchedGemmSoftmaxGemm<ALayout,
B0Layout,
......@@ -65,9 +84,10 @@ struct DeviceOperationInstanceFactory<
CDataType,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
PassThrough>;
MaskOutUpperTriangle>;
static auto GetInstances()
{
......@@ -78,11 +98,19 @@ struct DeviceOperationInstanceFactory<
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
is_same_v<B1Layout, Row> && is_same_v<CLayout, Row>)
{
if constexpr(MaskOutUpperTriangle)
{
add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
op_ptrs);
}
else
{
add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
op_ptrs);
}
}
}
return op_ptrs;
}
};
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute.hpp → library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
View file @ de37550f
......@@ -17,63 +17,89 @@ namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
F16,
F16,
F16,
F16,
ck::Tuple<>,
ck::Tuple<>,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
MaskingSpecialization::MaskOutUpperTriangle>>>&
instances);
void add_device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<Row,
Col,
Row,
CPermuteNumDims_G_M_O,
void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<
std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
F16,
F16,
F16,
F16,
ck::Tuple<>,
ck::Tuple<>,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough>>>& instances);
PassThrough,
MaskingSpecialization::MaskDisabled>>>&
instances);
template <typename ALayout,
typename B0Layout,
typename B1Layout,
typename CPermuteNumDims_G_M_Gemm1N,
typename ADataType,
template <typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType>
typename CDataType,
MaskingSpecialization MaskingSpec>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_Gemm1N,
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
ADataType,
B0DataType,
B1DataType,
CDataType,
ck::Tuple<>,
ck::Tuple<>,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough>>
PassThrough,
MaskingSpec>>
{
using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_Gemm1N,
using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
ADataType,
B0DataType,
B1DataType,
CDataType,
ck::Tuple<>,
ck::Tuple<>,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough>;
PassThrough,
MaskingSpec>;
static auto GetInstances()
{
......@@ -82,11 +108,14 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
is_same_v<B1Layout, Row> &&
is_same_v<CPermuteNumDims_G_M_Gemm1N, CPermuteNumDims_G_M_O>)
if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
{
add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
op_ptrs);
}
else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
{
add_device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
op_ptrs);
}
}
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ de37550f
......@@ -6,6 +6,7 @@ function(add_instance_library INSTANCE_NAME)
clang_tidy_check(${INSTANCE_NAME})
endfunction(add_instance_library INSTANCE_NAME)
file(GLOB dir_list LIST_DIRECTORIES true *)
set(CK_DEVICE_INSTANCES)
FOREACH(subdir_path ${dir_list})
......
library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @ de37550f
......@@ -36,10 +36,10 @@ using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_inst
//################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>, // failed validation on MI100
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
......
library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
View file @ de37550f
......@@ -36,10 +36,10 @@ using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_inst
//################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 4, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>, // TODO: to enable; can trigger compiler crash in mainline #9110 but not in #10738
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 4, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 4, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 4, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>, // TODO: to enable; can cause validation error on MI100
// DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 4, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>, // TODO: to enable; can cause validation error on MI100
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 4, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 4, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 4, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 4, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmGemm_Xdl_CShuffle< Row, Col, Col, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 4, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, 1, 2, S<1, 32, 1, 8>, 8>,
......
library/src/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute/CMakeLists.txt deleted 100644 → 0
View file @ cd517326
add_instance_library(device_batched_gemm_masking_scale_softmax_gemm_permute_instance
device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
)
library/src/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute/device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp deleted 100644 → 0
View file @ cd517326
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#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_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmPadded = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
using device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
std::tuple<
// clang-format off
// 2 of them are commented out because they trigger the clang-13 issue.
//##############################################| ALayout| B0Layout| B1Layout| CPermuteNumDims_G_M_O| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskOut|
//##############################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//##############################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
//DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
//DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, true>,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>
// clang-format on
>;
void add_device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<Row,
Col,
Row,
CPermuteNumDims_G_M_O,
F16,
F16,
F16,
F16,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @ de37550f
......@@ -24,11 +24,13 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmPadded = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
template <bool Masking>
using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
std::tuple<
// clang-format off
......@@ -36,24 +38,25 @@ using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, Masking>,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>
// clang-format on
>;
template <bool Masking>
using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances =
std::tuple<
// clang-format off
......@@ -61,12 +64,14 @@ using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 256, 128, 40, 64, 32, 4, 4, 2, 32, 32, 2, 4, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 256, 128, 40, 128, 32, 4, 4, 2, 32, 32, 2, 4, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 256, 40, 64, 32, 4, 4, 2, 32, 32, 1, 8, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 256, 40, 128, 32, 4, 4, 2, 32, 32, 1, 8, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 40, 64, 32, 4, 4, 2, 32, 32, 1, 4, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 40, 128, 32, 4, 4, 2, 32, 32, 1, 4, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 256, 128, 40, 64, 32, 4, 4, 2, 32, 32, 2, 4, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 256, 128, 40, 128, 32, 4, 4, 2, 32, 32, 2, 4, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
#if CK_WORKAROUND_DISABLE_BROKEN_ATTN_KERNEL_INSTANCE == 0
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 256, 40, 64, 32, 4, 4, 2, 32, 32, 1, 8, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
#endif
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 256, 40, 128, 32, 4, 4, 2, 32, 32, 1, 8, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 40, 64, 32, 4, 4, 2, 32, 32, 1, 4, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 40, 128, 32, 4, 4, 2, 32, 32, 1, 4, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, Masking>
// clang-format on
>;
......@@ -81,16 +86,45 @@ void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_g
F16,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
PassThrough>>>& instances)
false>>>& instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
false>{});
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances{});
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances<
false>{});
}
void add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
Col,
Row,
Row,
F16,
F16,
F16,
F16,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
true>>>& instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
true>{});
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances<
true>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt 0 → 100644
View file @ de37550f
add_instance_library(device_batched_gemm_softmax_gemm_permute_instance
device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
)
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp 0 → 100644
View file @ de37550f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#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_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmPadded = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
index_t NumDimO,
MaskingSpecialization MaskingSpec>
using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
std::tuple<
// clang-format off
// #############################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskingSpec|
// #############################################| | | | | | Type| Type| Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| |
// #############################################| | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| |
// #############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, MaskingSpec>,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec>
// clang-format on
>;
void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
F16,
F16,
F16,
F16,
ck::Tuple<>,
ck::Tuple<>,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
MaskingSpecialization::MaskOutUpperTriangle>>>&
instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
2,
1,
1,
1,
1,
MaskingSpecialization::MaskOutUpperTriangle>{});
}
void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<
std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
F16,
F16,
F16,
F16,
ck::Tuple<>,
ck::Tuple<>,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough,
MaskingSpecialization::MaskDisabled>>>&
instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
2,
1,
1,
1,
1,
MaskingSpecialization::MaskDisabled>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profile_batched_gemm_softmax_gemm_impl.hpp
View file @ de37550f
......@@ -29,7 +29,8 @@ template <typename ADataType,
typename ALayout,
typename B0Layout,
typename B1Layout,
typename CLayout>
typename CLayout,
bool MaskOutUpperTriangle>
bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
int init_method,
bool do_log,
......@@ -46,16 +47,18 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
int BatchStrideA = -1,
int BatchStrideB0 = -1,
int BatchStrideB1 = -1,
int BatchStrideC = -1)
int BatchStrideC = -1,
float alpha = 1.f)
{
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
using PassThrough = tensor_operation::element_wise::PassThrough;
using Scale = tensor_operation::element_wise::Scale;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = PassThrough;
using Acc0ElementOp = Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
using AccDataType = float;
......@@ -67,7 +70,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
AccDataType,
AElementOp,
B0ElementOp,
CElementOp>;
Acc0ElementOp>;
// Ref Softmax: fp32 in, various type out
using ReferenceSoftmaxInstance =
......@@ -185,7 +188,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
......@@ -201,7 +204,8 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp>;
CElementOp,
MaskOutUpperTriangle>;
// get device op instances
const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -214,10 +218,16 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, PassThrough{});
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, Scale{alpha});
ref_gemm0_invoker.Run(ref_gemm0_argument);
// mask out upper triangle
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(MaskOutUpperTriangle && idx[1] < idx[2])
self(idx) = -ck::NumericLimits<float>::Infinity();
});
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
......
profiler/include/profile_batched_gemm_masking_scale_softmax_gemm_permute_impl.hpp → profiler/include/profile_batched_gemm_softmax_gemm_permute_impl.hpp
View file @ de37550f
......@@ -7,10 +7,10 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
......@@ -22,15 +22,19 @@
namespace ck {
namespace profiler {
template <typename ADataType,
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
index_t NumDimO,
typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType,
typename ALayout,
typename B0Layout,
typename B1Layout,
typename CPermuteNumDims_G_M_O>
bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verification,
typename Acc0BiasesDataType,
typename Acc1BiasesDataType,
tensor_operation::device::MaskingSpecialization MaskingSpec>
bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
int init_method,
bool do_log,
bool time_kernel,
......@@ -40,18 +44,10 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
int O,
int G0,
int G1,
int StrideA = -1,
int StrideB0 = -1,
int StrideB1 = -1,
int BatchStrideA = -1,
int BatchStrideB0 = -1,
int BatchStrideB1 = -1,
float alpha = 1.f)
{
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
using PassThrough = tensor_operation::element_wise::PassThrough;
using Scale = tensor_operation::element_wise::Scale;
using AElementOp = PassThrough;
......@@ -60,6 +56,7 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
using AccDataType = float;
using tensor_operation::device::MaskingSpecialization;
// Ref Gemm0: various type in, fp32 out
using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -85,67 +82,33 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
bool pass = true;
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
// A layout [G0, M, G1, K]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
// B0 layout [G0, N, G1, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
// B1 layout [G0, N, G1, O]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
// C layout [G0, M, G1, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
const int BatchCount = G0 * G1;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_gs_ms_os_host_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
Tensor<CDataType> c_gs_ms_os_device_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
// Host verification: Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{BatchCount, M, O},
std::vector<int>{M * O, O, 1});
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
std::srand(1); // work around test flakiness
......@@ -157,38 +120,38 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
// or not. May want to try exact same approach as the GPU kernel in the host reference
// GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
// shrink the input value range as it is less likely to produce errors of around ~1e-3.
// a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
// b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
// b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
// a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
// b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
// b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
DeviceMem c_gs_ms_os_device_buf(sizeof(CDataType) *
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
......@@ -196,20 +159,23 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
using DeviceOp =
tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
using DeviceOp = tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
1,
1,
1,
1,
ADataType,
B0DataType,
B1DataType,
CDataType,
ck::Tuple<>,
ck::Tuple<>,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp>;
CElementOp,
MaskingSpec>;
// get device op instances
const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -219,6 +185,26 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
if(do_verification)
{
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g_m_k({BatchCount, M, K});
Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({BatchCount, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
......@@ -228,7 +214,7 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
// mask out upper triangle
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(idx[1] < idx[2])
if(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle && idx[1] < idx[2])
self(idx) = -ck::NumericLimits<float>::Infinity();
});
......@@ -265,23 +251,24 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_gs_ms_os_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
BatchCount,
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
{}, // std::array<void*, 1> p_acc0_biases;
{}, // std::array<void*, 1> p_acc1_biases;
a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
{}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
a_element_op,
b0_element_op,
acc0_element_op,
......@@ -319,18 +306,18 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
if(do_verification)
{
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
pass = pass & ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
LogRangeAsType<float>(std::cout << "a_gs_ms_ks: ", a_gs_ms_ks.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
LogRangeAsType<float>(std::cout << "b0_gs_ns_ks : ", b0_gs_ns_ks.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
LogRangeAsType<float>(std::cout << "b1_gs_os_ns : ", b1_gs_os_ns.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")
......
test/CMakeLists.txt
View file @ de37550f
......@@ -41,7 +41,7 @@ add_subdirectory(batched_gemm)
add_subdirectory(batched_gemm_reduce)
add_subdirectory(batched_gemm_gemm)
add_subdirectory(batched_gemm_softmax_gemm)
add_subdirectory(batched_gemm_masking_scale_softmax_gemm_permute)
add_subdirectory(batched_gemm_softmax_gemm_permute)
add_subdirectory(grouped_gemm)
add_subdirectory(reduce)
add_subdirectory(convnd_fwd)
......
test/batched_gemm_masking_scale_softmax_gemm_permute/CMakeLists.txt deleted 100644 → 0
View file @ cd517326
add_custom_target(test_batched_gemm_masking_scale_softmax_gemm_permute)
add_gtest_executable(test_batched_gemm_masking_scale_softmax_gemm_permute_fp16 test_batched_gemm_masking_scale_softmax_gemm_permute_fp16.cpp)
target_link_libraries(test_batched_gemm_masking_scale_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_masking_scale_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_masking_scale_softmax_gemm_permute test_batched_gemm_masking_scale_softmax_gemm_permute_fp16)
\ No newline at end of file
test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
View file @ de37550f
......@@ -9,9 +9,13 @@ class TestBatchedGemmSoftmaxGemmFP16 : public TestBatchedGemmSoftmaxGemm<Tuple>
{
};
using Masked = std::true_type;
using NoMask = std::false_type;
// clang-format off
using KernelTypes = ::testing::Types<
std::tuple<F16, F16, F16, F16, Row, Col, Row, Row>
std::tuple<F16, F16, F16, F16, Row, Col, Row, Row, NoMask>,
std::tuple<F16, F16, F16, F16, Row, Col, Row, Row, Masked>
>;
// clang-format on
......@@ -120,7 +124,6 @@ TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16_IrregularK)
using ck::tensor_operation::device::GemmSpecialization;
// TODO: enable KPadding tests when it is implemented
TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMatch)
{
int P = 120; // requires padding
......@@ -152,12 +155,12 @@ TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMismatch)
// IsSupported(M, N, K, O)
// clang-format off
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
// Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
// Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
// clang-format on
}
......@@ -169,6 +172,5 @@ TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, AdhocTest)
{1020, 1020, 64, 128, 24},
{576, 576, 64, 64, 24},
};
this->bench_ = true;
this->Run();
}
test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
View file @ de37550f
......@@ -28,6 +28,7 @@ struct TestBatchedGemmSoftmaxGemm : public ::testing::Test
using B0Layout = std::tuple_element_t<5, Tuple>;
using B1Layout = std::tuple_element_t<6, Tuple>;
using CLayout = std::tuple_element_t<7, Tuple>;
using MaskingType = std::tuple_element_t<8, Tuple>;
std::vector<std::vector<int>> lengths_ = {{256, 256, 64, 64, 4},
{256, 256, 128, 128, 4},
......@@ -54,7 +55,8 @@ struct TestBatchedGemmSoftmaxGemm : public ::testing::Test
ALayout,
B0Layout,
B1Layout,
CLayout>(
CLayout,
MaskingType::value>(
verify_, 1, false, bench_, M, N, K, O, BatchCount);
EXPECT_TRUE(pass);
......
test/batched_gemm_softmax_gemm_permute/CMakeLists.txt 0 → 100644
View file @ de37550f
add_custom_target(test_batched_gemm_softmax_gemm_permute)
add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp)
target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
\ No newline at end of file
test/batched_gemm_masking_scale_softmax_gemm_permute/test_batched_gemm_masking_scale_softmax_gemm_permute_fp16.cpp → test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
View file @ de37550f
......@@ -2,7 +2,7 @@
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "test_batched_gemm_masking_scale_softmax_gemm_permute_util.hpp"
#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
template <typename Tuple>
class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16
......@@ -10,13 +10,18 @@ class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16
{
};
using I1_t = ck::Number<1>;
using I2_t = ck::Number<2>;
using MaskDisabled_t =
ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
using MaskOutUpperTriangle_t =
ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
// clang-format off
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
using KernelTypes = ::testing::Types<
std::tuple<F16, F16, F16, F16, Row, Col, Row, CPermuteNumDims_G_M_O>
std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, MaskDisabled_t>,
std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, MaskOutUpperTriangle_t>
>;
// clang-format on
......@@ -91,7 +96,7 @@ TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddO)
this->Run();
}
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Bench_FP16_IrregularK)
TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16_IrregularK)
{
this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
{256, 64, 160, 64, 1, 16},
......@@ -125,7 +130,6 @@ TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP1
using ck::tensor_operation::device::GemmSpecialization;
// TODO: enable KPadding tests when it is implemented
TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
{
int P = 120; // requires padding
......@@ -133,22 +137,22 @@ TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationS
// IsSupported(M, N, K, O)
// clang-format off
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
// clang-format on
}
......@@ -156,13 +160,13 @@ TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationS
{
// IsSupported(M, N, K, O)
// clang-format off
EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
// Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
// Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
// EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
// clang-format on
}
......@@ -174,6 +178,5 @@ TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, AdhocTest)
{1020, 1020, 64, 128, 4, 6},
{576, 576, 64, 64, 4, 6},
};
this->bench_ = true;
this->Run();
}
test/batched_gemm_masking_scale_softmax_gemm_permute/test_batched_gemm_masking_scale_softmax_gemm_permute_util.hpp → test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
View file @ de37550f
......@@ -4,10 +4,14 @@
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "profiler/include/profile_batched_gemm_masking_scale_softmax_gemm_permute_impl.hpp"
#include "profiler/include/profile_batched_gemm_softmax_gemm_permute_impl.hpp"
using ck::tensor_operation::device::GemmSpecialization;
using ck::tensor_operation::device::MaskingSpecialization;
using ck::tensor_operation::device::TensorSpecialization;
template <ck::index_t N>
using I = ck::Number<N>;
......@@ -20,14 +24,18 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
template <typename Tuple>
struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
{
using ADataType = std::tuple_element_t<0, Tuple>;
using B0DataType = std::tuple_element_t<1, Tuple>;
using B1DataType = std::tuple_element_t<2, Tuple>;
using CDataType = std::tuple_element_t<3, Tuple>;
using ALayout = std::tuple_element_t<4, Tuple>;
using B0Layout = std::tuple_element_t<5, Tuple>;
using B1Layout = std::tuple_element_t<6, Tuple>;
using CPermuteNumDims_G_M_O = std::tuple_element_t<7, Tuple>;
using NumDimGType = std::tuple_element_t<0, Tuple>;
using NumDimMType = std::tuple_element_t<1, Tuple>;
using NumDimNType = std::tuple_element_t<2, Tuple>;
using NumDimKType = std::tuple_element_t<3, Tuple>;
using NumDimOType = std::tuple_element_t<4, Tuple>;
using ADataType = std::tuple_element_t<5, Tuple>;
using B0DataType = std::tuple_element_t<6, Tuple>;
using B1DataType = std::tuple_element_t<7, Tuple>;
using CDataType = std::tuple_element_t<8, Tuple>;
using Acc0BiasDataType = std::tuple_element_t<9, Tuple>;
using Acc1BiasDataType = std::tuple_element_t<10, Tuple>;
using MaskingType = std::tuple_element_t<11, Tuple>;
std::vector<std::vector<int>> lengths_ = {
{256, 256, 64, 64, 6, 4},
......@@ -42,15 +50,20 @@ struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
void RunSingle(int M, int N, int K, int O, int G0, int G1)
{
bool pass = ck::profiler::profile_batched_gemm_masking_scale_softmax_gemm_permute_impl<
bool pass =
ck::profiler::profile_batched_gemm_softmax_gemm_permute_impl<NumDimGType::value,
NumDimMType::value,
NumDimNType::value,
NumDimKType::value,
NumDimOType::value,
ADataType,
B0DataType,
B1DataType,
CDataType,
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O>(verify_, 1, false, bench_, M, N, K, O, G0, G1);
ck::Tuple<>,
ck::Tuple<>,
MaskingType::value>(
verify_, 1, false, bench_, M, N, K, O, G0, G1);
EXPECT_TRUE(pass);
}
......@@ -72,19 +85,13 @@ struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
};
template <GemmSpecialization GemmSpec>
struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
using ADataType = F16;
using B0DataType = F16;
......@@ -103,14 +110,17 @@ struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
using DeviceGemmGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
2,
1,
1,
1,
1,
ADataType,
B0DataType,
B1DataType,
CDataType,
ck::Tuple<>,
ck::Tuple<>,
AccDataType,
CShuffleDataType,
AElementOp,
......@@ -119,6 +129,10 @@ struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecialization::Default, // ATensorSpec
TensorSpecialization::Default, // B0TensorSpec
TensorSpecialization::Default, // B1TensorSpec
TensorSpecialization::Default, // CTensorSpec
1,
256,
128, // MPerBlock
......@@ -159,29 +173,48 @@ struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
true>; // Masking
MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
bool IsSupported(int M, int N, int K, int O)
{
const int G0 = 1, G1 = 1;
// A layout [G0, M, G1, K]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
// B0 layout [G0, N, G1, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
// B1 layout [G0, N, G1, O]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
// C layout [G0, M, G1, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
auto gemm = DeviceGemmGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
static_cast<B0DataType*>(nullptr),
static_cast<B1DataType*>(nullptr),
static_cast<CDataType*>(nullptr),
M,
N,
K,
O,
0, // BatchCount
{0, 0, M, O}, // gs ms ns lengths
{0, O, 0, 1}, // gs ms ns strides
0, // StrideA
0, // StrideB0
0, // StrideB1
0, // BatchStrideA
0, // BatchStrideB0
0, // BatchStrideB1
{}, // p_acc0_biases
{}, // p_acc1_biases
a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
{}, // acc0_biases_gs_ms_ns_lengths
{}, // acc0_biases_gs_ms_ns_strides
{}, // acc1_biases_gs_ms_os_lengths
{}, // acc1_biases_gs_ms_os_strides
PassThrough{}, // a_element_op
PassThrough{}, // b0_element_op
Scale{1.f}, // acc0_element_op
......
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