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
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1655 additions and 1725 deletions
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @ de37550f
......@@ -2,9 +2,11 @@ add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_custom_target(example_gemm_scale_softmax_gemm)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ de37550f
......@@ -33,9 +33,6 @@ using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
......@@ -44,13 +41,14 @@ using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
......@@ -59,17 +57,27 @@ using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
......@@ -78,6 +86,10 @@ using DeviceGemmInstance =
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
......@@ -118,7 +130,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
true>; // MaskOutUpperTriangle
MaskingSpec>; // MaskingSpecialization
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -142,268 +154,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
B1ElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 512;
ck::index_t N = 512;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t StrideA = -1;
ck::index_t StrideB0 = -1;
ck::index_t StrideB1 = -1;
ck::index_t BatchStrideA = -1;
ck::index_t BatchStrideB0 = -1;
ck::index_t BatchStrideB1 = -1;
float alpha = 1;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
exit(0);
}
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;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
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;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
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()));
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;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
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});
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});
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>{});
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>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
DeviceMem c_gs_ms_os_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());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument =
gemm.MakeArgument(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,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
if(do_verification)
{
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
// 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});
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, acc0_element_op);
// gemm 0
ref_gemm0_invoker.Run(ref_gemm0_argument);
// mask out upper triangle
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(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});
// softmax
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
// gemm1
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
? 0
: 1;
}
#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
return 0;
}
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ de37550f
......@@ -33,9 +33,6 @@ using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
......@@ -44,13 +41,14 @@ using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
......@@ -59,17 +57,27 @@ using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
......@@ -78,6 +86,10 @@ using DeviceGemmInstance =
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
......@@ -118,7 +130,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
false>; // MaskOutUpperTriangle
MaskingSpec>; // MaskingSpecialization
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -142,258 +154,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
B1ElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 120;
ck::index_t N = 1000;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t StrideA = -1;
ck::index_t StrideB0 = -1;
ck::index_t StrideB1 = -1;
ck::index_t BatchStrideA = -1;
ck::index_t BatchStrideB0 = -1;
ck::index_t BatchStrideB1 = -1;
float alpha = 1;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
exit(0);
}
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;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
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;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
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()));
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;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
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});
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});
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>{});
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>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
DeviceMem c_gs_ms_os_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());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument =
gemm.MakeArgument(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,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
if(do_verification)
{
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
// 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});
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, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
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});
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
? 0
: 1;
}
#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
return 0;
}
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp 0 → 100644
View file @ de37550f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec>; // MaskingSpecialization
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
AccDataType,
AccDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
View file @ de37550f
......@@ -33,9 +33,6 @@ using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
......@@ -44,13 +41,14 @@ using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CPermuteNumDims_G_M_O =
S<1, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_M_O
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
......@@ -59,17 +57,27 @@ using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
......@@ -78,6 +86,10 @@ using DeviceGemmInstance =
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
......@@ -118,7 +130,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
false>;
MaskingSpec>; // MaskingSpecialization
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -142,303 +154,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
B1ElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
exit(0);
}
float alpha = 1; // scaling after 1st gemm
std::size_t group_count = 13;
// Problem descs
std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
std::vector<const void*> p_a;
std::vector<const void*> p_b0;
std::vector<const void*> p_b1;
std::vector<void*> p_c;
for(std::size_t i = 0; i < group_count; i++)
{
int M = 128 * (rand() % 8 + 1);
int N = 128 * (rand() % 8 + 1);
int K = 40;
int O = 40 * (rand() % 2 + 1);
int Batch = rand() % 8 + 1;
const int StrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int StrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int StrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
const int BatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int BatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int BatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
std::vector<ck::index_t> c_gs_ms_os_lengths{Batch, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{O, Batch * O, 1};
problem_descs.push_back({M,
N,
K,
O,
Batch,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
c_gs_ms_os_lengths,
c_gs_ms_os_strides});
}
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}));
}
};
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<CDataType>> c_tensors;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device;
std::vector<DeviceMemPtr> b0_tensors_device;
std::vector<DeviceMemPtr> b1_tensors_device;
std::vector<DeviceMemPtr> c_tensors_device;
std::size_t flop = 0, num_byte = 0;
std::cout << "group count " << group_count << ". printing first 4 groups\n";
for(std::size_t i = 0; i < group_count; i++)
{
const auto& M = problem_descs[i].M;
const auto& N = problem_descs[i].N;
const auto& K = problem_descs[i].K;
const auto& O = problem_descs[i].O;
const auto& Batch = problem_descs[i].Batch;
const auto& StrideA = problem_descs[i].StrideA;
const auto& StrideB0 = problem_descs[i].StrideB0;
const auto& StrideB1 = problem_descs[i].StrideB1;
const auto& BatchStrideA = problem_descs[i].BatchStrideA;
const auto& BatchStrideB0 = problem_descs[i].BatchStrideB0;
const auto& BatchStrideB1 = problem_descs[i].BatchStrideB1;
const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(Batch, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(Batch, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(Batch, N, O, StrideB1, BatchStrideB1, B1Layout{}));
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()));
flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
Batch;
if(i < 4)
{
std::cout << "a_g_m_k[" << i << "]: " << a_g_m_k.mDesc << ", "
<< "b0_g_k_n[" << i << "]: " << b0_g_k_n.mDesc << ", "
<< "b1_g_n_o[" << i << "]: " << b1_g_n_o.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
}
switch(init_method)
{
case 0: break;
case 1:
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});
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});
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>{});
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_tensors.push_back(a_g_m_k);
b0_tensors.push_back(b0_g_k_n);
b1_tensors.push_back(b1_g_n_o);
c_tensors.push_back(c_gs_ms_os_device_result);
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize()));
b0_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize()));
b1_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_g_m_k.mData.data());
b0_tensors_device[i]->ToDevice(b0_g_k_n.mData.data());
b1_tensors_device[i]->ToDevice(b1_g_n_o.mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(p_a,
p_b0,
p_b1,
p_c,
problem_descs,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
// specify workspace for problem_desc
DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
for(std::size_t i = 0; i < group_count; i++)
{
const auto& M = problem_descs[i].M;
const auto& N = problem_descs[i].N;
const auto& O = problem_descs[i].O;
const auto& Batch = problem_descs[i].Batch;
const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
const auto& a_g_m_k = a_tensors[i];
const auto& b0_g_k_n = b0_tensors[i];
const auto& b1_g_n_o = b1_tensors[i];
auto& c_gs_ms_os_device_result = c_tensors[i];
auto& c_gs_ms_os_device_buf = *c_tensors_device[i];
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()));
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
// Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{Batch, M, O},
std::vector<int>{M * O, O, 1});
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_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// Note: in this example, we merely permute the dimensions by changing underlying
// strides so we simply access data as-is
c_gs_ms_os_host_result.ForEach(
[&](auto& self, auto idx) { self(idx) = c_g_m_o_host_result(idx); });
bool pass_ =
ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
pass &= pass_;
}
}
#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
return pass ? 0 : 1;
}
int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc 0 → 100644
View file @ de37550f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 120;
ck::index_t N = 1000;
ck::index_t K = 64;
ck::index_t O = 128;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
float alpha = 1;
bool input_permute = false;
bool output_permute = true;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 13)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
alpha = std::stof(argv[10]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
exit(0);
}
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, 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 =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, 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 =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
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;
switch(init_method)
{
case 0: break;
case 1:
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_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_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_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
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_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{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
// TODO ANT: replace array with vector?
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(
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,
{}, // 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,
b1_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
ck::index_t BatchCount = G0 * G1;
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
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);
});
// gemm 0
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, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = DeviceGemmInstance::C0MatrixMask(N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
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});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
? 0
: 1;
}
return 0;
}
example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc 0 → 100644
View file @ de37550f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool input_permute = false;
bool output_permute = true;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
input_permute = std::stoi(argv[4]);
output_permute = std::stoi(argv[5]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 5: input / output permute\n");
exit(0);
}
float alpha = 1; // scaling after 1st gemm
std::size_t group_count = 7;
// Problem descs
std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
std::vector<const void*> p_a;
std::vector<const void*> p_b0;
std::vector<const void*> p_b1;
std::vector<void*> p_c;
std::vector<std::vector<int>> g0_g1_m_n_k_o;
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<CDataType>> c_tensors;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device;
std::vector<DeviceMemPtr> b0_tensors_device;
std::vector<DeviceMemPtr> b1_tensors_device;
std::vector<DeviceMemPtr> c_tensors_device;
std::size_t flop = 0, num_byte = 0;
std::cout << "group count " << group_count << ". printing first 4 groups\n";
for(std::size_t i = 0; i < group_count; i++)
{
int M = 128 * (rand() % 8 + 1);
int N = 128 * (rand() % 8 + 1);
int K = 40;
int O = 40 * (rand() % 2 + 1);
int G0 = rand() % 3 + 1;
int G1 = rand() % 5 + 1;
g0_g1_m_n_k_o.push_back({G0, G1, M, N, K, O});
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, 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 =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, 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 =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
problem_descs.push_back({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
// C_m_o = A_m_k * B0_k_n * B1_n_o
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_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
int Batch = G0 * G1;
flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
Batch;
if(i < 4)
{
std::cout << "a_gs_ms_ks[" << i << "]: " << a_gs_ms_ks.mDesc << ", "
<< "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
<< "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
}
switch(init_method)
{
case 0: break;
case 1:
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_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_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_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
a_tensors.push_back(a_gs_ms_ks);
b0_tensors.push_back(b0_gs_ns_ks);
b1_tensors.push_back(b1_gs_os_ns);
c_tensors.push_back(c_gs_ms_os_device_result);
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize()));
b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize()));
b1_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_gs_ms_ks.mData.data());
b0_tensors_device[i]->ToDevice(b0_gs_ns_ks.mData.data());
b1_tensors_device[i]->ToDevice(b1_gs_os_ns.mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(p_a,
p_b0,
p_b1,
p_c,
{}, // p_acc0_biases
{}, // p_acc1_biases
problem_descs,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
// specify workspace for problem_desc
DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
for(std::size_t i = 0; i < group_count; i++)
{
const int& G0 = g0_g1_m_n_k_o[i][0];
const int& G1 = g0_g1_m_n_k_o[i][1];
const int& M = g0_g1_m_n_k_o[i][2];
const int& N = g0_g1_m_n_k_o[i][3];
const int& K = g0_g1_m_n_k_o[i][4];
const int& O = g0_g1_m_n_k_o[i][5];
const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
const auto& a_gs_ms_ks = a_tensors[i];
const auto& b0_gs_ns_ks = b0_tensors[i];
const auto& b1_gs_os_ns = b1_tensors[i];
auto& c_gs_ms_os_device_result = c_tensors[i];
auto& c_gs_ms_os_device_buf = *c_tensors_device[i];
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
Tensor<ADataType> a_g_m_k({G0 * G1, M, K});
Tensor<B0DataType> b0_g_k_n({G0 * G1, K, N});
Tensor<B1DataType> b1_g_n_o({G0 * G1, N, O});
Tensor<AccDataType> acc0_g_m_n({G0 * G1, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({G0 * G1, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({G0 * G1, M, O}); // scratch object after gemm1
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
// 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);
});
// gemm 0
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, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = DeviceGemmInstance::C0MatrixMask(N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
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});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm 1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
bool pass_ =
ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
pass &= pass_;
}
}
return pass ? 0 : 1;
}
include/ck/ck.hpp
View file @ de37550f
......@@ -159,6 +159,11 @@
// tuning parameter
#define CK_WORKAROUND_SWDEV_325164 0
// workaround: disable broken fused attention kernel instance that does not pass validation
// issue found on mi100/#10738 combo when irregular KPerBlock attention kernel has acc0 scaling
// enabled
#define CK_WORKAROUND_DISABLE_BROKEN_ATTN_KERNEL_INSTANCE 1
namespace ck {
enum struct InMemoryDataOperationEnum
......
include/ck/tensor_description/tensor_space_filling_curve.hpp
View file @ de37550f
......@@ -14,7 +14,8 @@ namespace ck {
template <typename TensorLengths,
typename DimAccessOrder,
typename ScalarsPerAccess> // # of scalars per access in each dimension
typename ScalarsPerAccess,
bool SnakeCurved = true> // # of scalars per access in each dimension
struct SpaceFillingCurve
{
static constexpr index_t nDim = TensorLengths::Size();
......@@ -136,9 +137,10 @@ struct SpaceFillingCurve
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto idim) {
ordered_idx(idim) = forward_sweep[idim] ? ordered_access_idx[idim]
: ordered_access_lengths[idim] - 1 -
ordered_access_idx[idim];
ordered_idx(idim) =
!SnakeCurved || forward_sweep[idim]
? ordered_access_idx[idim]
: ordered_access_lengths[idim] - 1 - ordered_access_idx[idim];
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View file @ de37550f
......@@ -151,6 +151,27 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
return make_tuple(c_thread_m, c_thread_n);
}
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
__device__ static auto
CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
return make_tuple(Number<m0>{},
Number<n0>{},
waveId_m,
waveId_n,
blk_idx[I0],
blk_idx[I1],
blk_idx[I2],
blk_idx[I3]);
}
__host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1()
{
static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
......@@ -724,6 +745,21 @@ struct BlockwiseGemmXdlops_v2
return make_tuple(c_thread_m, c_thread_n);
}
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
__device__ static auto
CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
return make_tuple(
m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
}
using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
__host__ __device__ BlockwiseGemmXdlops_v2(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
View file @ de37550f
......@@ -24,7 +24,8 @@ template <typename ALayout,
typename B0ElementwiseOperation,
typename Acc0ElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
bool MaskOutUpperTriangle> // TODO: enum for mask type
struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
View file @ de37550f
......@@ -7,49 +7,60 @@
#include <vector>
#include "device_base.hpp"
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename B0Layout,
typename B1Layout,
typename CPermuteNumDims_G_M_Gemm1N, // Sequence<>
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 Acc0BiasDataType,
typename Acc1BiasDataType,
typename AElementwiseOperation,
typename B0ElementwiseOperation,
typename Acc0ElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
MaskingSpecialization MaskingSpec>
struct DeviceBatchedGemmSoftmaxGemmPermute : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b0,
const void* p_b1,
void* p_c,
ck::index_t M,
ck::index_t N,
ck::index_t K,
ck::index_t O,
ck::index_t Batch,
std::vector<index_t> c_gs_ms_os_lengths,
std::vector<index_t> c_gs_ms_os_strides,
ck::index_t StrideA,
ck::index_t StrideB0,
ck::index_t StrideB1,
ck::index_t BatchStrideA,
ck::index_t BatchStrideB0,
ck::index_t BatchStrideB1,
AElementwiseOperation a_element_op,
B0ElementwiseOperation b0_element_op,
Acc0ElementwiseOperation acc0_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op) = 0;
static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
const void* p_a,
const void* p_b0,
const void* p_b1,
void* p_c,
const std::array<void*, NumAcc0Bias> p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths,
const std::vector<index_t>& b_gs_ns_ks_strides,
const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
const std::vector<index_t>& c_gs_ms_gemm1ns_lengths, // c_gs_ms_os_lengths
const std::vector<index_t>& c_gs_ms_gemm1ns_strides, // c_gs_ms_os_strides
const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<index_t>, NumAcc1Bias>
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<index_t>, NumAcc1Bias>
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op,
B0ElementwiseOperation b0_element_op,
Acc0ElementwiseOperation acc0_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
View file @ de37550f
......@@ -7,46 +7,50 @@
#include <vector>
#include "device_base.hpp"
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename B0Layout,
typename B1Layout,
typename CPermuteNumDims_G_M_Gemm1N, // Sequence<>
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 Acc0BiasDataType,
typename Acc1BiasDataType,
typename AElementwiseOperation,
typename B0ElementwiseOperation,
typename Acc0ElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
MaskingSpecialization MaskingSpec>
struct DeviceGroupedGemmSoftmaxGemmPermute : public BaseOperator
{
struct ProblemDesc
{
// Overall problem shape
index_t M;
index_t N;
index_t K;
index_t O;
index_t Batch;
std::vector<index_t> a_gs_ms_ks_lengths;
std::vector<index_t> a_gs_ms_ks_strides;
// Stride for A/B0/B1; layout determined by template args
index_t StrideA;
index_t StrideB0;
index_t StrideB1;
index_t BatchStrideA;
index_t BatchStrideB0;
index_t BatchStrideB1;
std::vector<index_t> b0_gs_ns_ks_lengths;
std::vector<index_t> b0_gs_ns_ks_strides;
std::vector<index_t> b1_gs_os_ns_lengths;
std::vector<index_t> b1_gs_os_ns_strides;
// Lengths and strides for output C
std::vector<index_t> c_gs_ms_os_lengths;
std::vector<index_t> c_gs_ms_os_strides;
std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_lengths;
std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_strides;
std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_lengths;
std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_strides;
};
virtual std::unique_ptr<BaseArgument>
......@@ -54,6 +58,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute : public BaseOperator
std::vector<const void*> p_b0_vec,
std::vector<const void*> p_b1_vec,
std::vector<void*> p_c_vec,
std::vector<std::vector<const void*>> p_acc0_biases_vec,
std::vector<std::vector<const void*>> p_acc1_biases_vec,
std::vector<ProblemDesc> problem_desc_vec,
AElementwiseOperation a_element_op,
B0ElementwiseOperation b0_element_op,
......
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
View file @ de37550f
......@@ -130,8 +130,11 @@ namespace device {
// D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
// E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
// FIXME: TensorSpecialization::Packed specialization does not cover all packed tensor cases, it
// merely degenerates into TensorSpecialization::Default with NumDimG/M/N/K = 1
// NOTE: TensorSpecialization::Packed specialized tensor is "packed" in a sense that each inner
// dimension in a dimension group (eg [G0, G1] in Gs, [M0, M1, M2] in Ms, etc.) are contiguous and
// ordered. Not in a sense that the tensor [G0, G1, ..., M0, M1, ..., N0, N1...] can be permuted
// while still being a contiguous, unpadded tensor. In other words, it merely degenerates into
// TensorSpecialization::Default with NumDimG/M/N/K = 1
//
// Detail- Packed tensor satisfies
// stride_0 = 1
......@@ -147,7 +150,7 @@ namespace device {
// essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
//
// Might need to expose dimension order to the interface to fully support
// TensorSpecialization::Packed.
// TensorSpecialization::Packed in a traditional sense of "packed" tensor
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
View file @ de37550f
......@@ -12,6 +12,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
#include "ck/host_utility/device_prop.hpp"
......@@ -196,7 +197,8 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation>
CElementwiseOperation,
MaskOutUpperTriangle>
{
using DeviceOp = DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle;
......@@ -315,29 +317,6 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
}
// to track the points which need to be set to -inf on C0
// Note: no need to reset M padding value, because they will not be stored out.
struct C0MatrixMask
{
C0MatrixMask(index_t NRaw) : NRaw_(NRaw) {}
__host__ __device__ bool IsUpperTriangle(index_t m, index_t n) const { return n > m; }
__host__ __device__ bool IsNOutOfBound(/*index_t m, */ index_t n) const
{
return n >= NRaw_;
}
__host__ __device__ bool IsMaskedElement(index_t m, index_t n) const
{
return IsUpperTriangle(m, n) || IsNOutOfBound(n);
}
private:
// index_t MRaw_;
index_t NRaw_;
};
struct ComputeBasePtrOfStridedBatch
{
ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
......@@ -383,6 +362,10 @@ struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
using C0MatrixMask = conditional_t<MaskOutUpperTriangle,
C0MatrixMask_impl<MaskOutUpperTrianglePredicate>,
C0MatrixMask_impl<MaskDisabledPredicate>>;
// GridwiseGemm
using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
ADataType, // TODO: distinguish A/B datatype
......
include/ck/tensor_operation/gpu/device/masking_specialization.hpp 0 → 100644
View file @ de37550f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
enum struct MaskingSpecialization
{
MaskDisabled,
MaskOutUpperTriangle
};
inline std::string getMaskingSpecializationString(const MaskingSpecialization& s)
{
switch(s)
{
case MaskingSpecialization::MaskDisabled: return "MaskDisabled";
case MaskingSpecialization::MaskOutUpperTriangle: return "MaskOutUpperTriangle";
default: return "Unrecognized specialization!";
}
}
struct MaskDisabledPredicate
{
__host__ __device__ constexpr bool operator()(index_t /*m*/, index_t /*n*/) const
{
return false;
};
__host__ __device__ constexpr bool
IsTileSkippable(index_t /*m*/, index_t /*n*/, index_t /*m_tile*/, index_t /*n_tile*/) const
{
return false;
}
};
struct MaskOutUpperTrianglePredicate
{
__host__ __device__ constexpr bool operator()(index_t m, index_t n) const { return n > m; }
__host__ __device__ constexpr bool
IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t /*n_tile*/) const
{
return operator()(m + m_tile - 1, n);
}
};
// to track the points which need to be set to -inf on C0
// Note: no need to reset M padding value, because they will not be stored out.
template <typename MaskOutPredicate>
struct C0MatrixMask_impl
{
C0MatrixMask_impl(index_t NRaw) : NRaw_(NRaw), predicate_(MaskOutPredicate{}) {}
__host__ __device__ constexpr bool IsNOutOfBound(/*index_t m, */ index_t n) const
{
return n >= NRaw_;
}
__host__ __device__ constexpr bool IsMaskedElement(index_t m, index_t n) const
{
return predicate_(m, n) || IsNOutOfBound(n);
}
__host__ __device__ constexpr bool
IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t n_tile) const
{
return predicate_.IsTileSkippable(m, n, m_tile, n_tile);
}
private:
// index_t MRaw_;
index_t NRaw_;
MaskOutPredicate predicate_;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
View file @ de37550f
......@@ -336,36 +336,6 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
};
template <bool Pred>
struct ElementOpPredicatedResetNaNToMinusInf;
template <>
struct ElementOpPredicatedResetNaNToMinusInf<true>
{
template <typename ElementOp, typename OutT, typename InT>
__host__ __device__ void Run(OutT& y, const ElementOp& op, const InT& x)
{
if(ck::math::isnan(x))
{
y = -ck::NumericLimits<float>::Infinity();
}
else
{
op(y, x);
}
}
};
template <>
struct ElementOpPredicatedResetNaNToMinusInf<false>
{
template <typename ElementOp, typename OutT, typename InT>
__host__ __device__ void Run(OutT& y, const ElementOp& op, const InT& x)
{
op(y, x);
}
};
template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
......@@ -406,11 +376,11 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
return;
}
// HACK: this force m/n_block_data_idx_on_grid into SGPR
// HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
const index_t n_block_data_idx_on_grid =
const index_t gemm1_n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
// A matrix in LDS memory, dst of blockwise copy
......@@ -627,7 +597,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
true, // DstResetCoord
NumGemmKPrefetchStage>(
b1_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
make_multi_index(0, gemm1_n_block_data_idx_on_grid, 0),
b1_element_op,
b1_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
......@@ -745,29 +715,16 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
running_max = NumericLimits<FloatGemmAcc>::Lowest();
running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
// decoder lower triangular mask
const auto thread_cluster_idx = threadid_to_m_n_thread_cluster_adaptor.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_n_cluster_id = thread_cluster_idx[I1];
const index_t MPerRepeat = MPerBlock / MXdlPerWave;
const index_t NPerRepeat = NPerBlock / NXdlPerWave;
const index_t mstart = m_block_data_idx_on_grid + thread_m_cluster_id;
// gemm1 K loop
index_t gemm1_k_block_outer_index = 0;
do
{
if constexpr(MaskOutUpperTriangle)
auto n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
if(c0_matrix_mask.IsTileSkippable(
m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
{
auto gemm0_n_block_idx =
__builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
if(c0_matrix_mask.IsUpperTriangle(m_block_data_idx_on_grid, gemm0_n_block_idx) &&
c0_matrix_mask.IsUpperTriangle(m_block_data_idx_on_grid + MPerBlock - 1,
gemm0_n_block_idx))
{
continue;
}
continue;
}
// gemm0
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
......@@ -789,60 +746,58 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
// do MNK padding or upper triangular masking
if constexpr(MaskOutUpperTriangle || PadN)
{
const index_t nstart = gemm1_k_block_outer_index * NPerBlock;
static_for<0, m0, 1>{}([&](auto m0_i) {
const index_t m_global = mstart + m0_i * MPerRepeat;
const index_t acc_idx_m0 = m0_i * n0 * n2 * n4;
static_for<0, n0, 1>{}([&](auto n0_i) {
// constexpr auto nrepeat_i = n0_i * NPerRepeat;
// const index_t nstartxdl = nstart + nrepeat_i;
const index_t nstartxdl = nstart + n0_i * NPerRepeat;
const index_t acc_idx_n0 = acc_idx_m0 + n0_i * n2 * n4;
static_for<0, n2, 1>{}([&](auto n2_i) {
const index_t nstartgroup =
nstartxdl + thread_n_cluster_id * n4 + n2_i * AccN3 * n4;
const index_t acc_idx_n2 = acc_idx_n0 + n2_i * n4;
static_for<0, n4, 1>{}([&](auto n4_i) {
const index_t n_global = nstartgroup + n4_i;
const auto acc_offset = Number<acc_idx_n2 + n4_i>{};
if constexpr(MaskOutUpperTriangle)
{
if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
{
acc_thread_buf(acc_offset) =
-ck::NumericLimits<float>::Infinity();
}
else
{
acc_element_op(acc_thread_buf(acc_offset),
acc_thread_buf[acc_offset]);
}
}
else
{
// ignore m_global;
if(c0_matrix_mask.IsNOutOfBound(n_global))
{
acc_thread_buf(acc_offset) =
-ck::NumericLimits<float>::Infinity();
}
else
{
acc_element_op(acc_thread_buf(acc_offset),
acc_thread_buf[acc_offset]);
}
}
});
});
});
// 8d thread_desc in thread scope
constexpr auto c_thread_lengths =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
// 8d block_desc in block scope
constexpr auto c_block_lengths =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
constexpr auto M0 = c_block_lengths[I0];
constexpr auto N0 = c_block_lengths[I1];
constexpr auto M1 = c_block_lengths[I2];
constexpr auto N1 = c_block_lengths[I3];
constexpr auto M2 = c_block_lengths[I4];
constexpr auto N2 = c_block_lengths[I5];
constexpr auto N3 = c_block_lengths[I6];
constexpr auto N4 = c_block_lengths[I7];
// works like multi-dimension static_for (static_ford), but provides both the linear
// index as well as n-d index
using Acc0TileIterator = SpaceFillingCurve<
decltype(c_thread_lengths),
typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
false>; // SnakeCurved
auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
auto m_local =
block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
auto n_local =
block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
auto m_global = m_local + m_block_data_idx_on_grid;
auto n_global = n_local + n_block_data_idx_on_grid;
if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
{
acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
}
else
{
acc_element_op(acc_thread_buf(i), acc_thread_buf[i]);
}
});
}
else
{
static_for<0, acc_thread_buf.Size(), 1>{}(
[&](auto i) { acc_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
}
block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ de37550f
......@@ -593,7 +593,8 @@ struct XdlopsGemm
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
using CIndex = MultiIndex<2>;
using CIndex = MultiIndex<2>;
using CIndex4D = MultiIndex<4>;
__device__ static constexpr index_t GetNumBlks() { return mfma_instr.num_output_blks; }
......@@ -822,6 +823,16 @@ struct XdlopsGemm
return TransposeC ? CIndex{n_offset, m_offset} : CIndex{m_offset, n_offset};
}
__device__ static CIndex4D GetBeginOfThreadBlk4D(index_t /* xdlops_i */, index_t /* blk_i */)
{
const auto blk_idx = GetBlkIdx();
const auto blk_id = blk_idx[I0];
const auto blk_td = blk_idx[I1];
return TransposeC ? CIndex4D{blk_td, I0, blk_id, I0} : CIndex4D{I0, blk_id, I0, blk_td};
}
static constexpr auto mfma = MfmaSelector<base_type, MPerXdlops, NPerXdlops>{};
static constexpr auto mfma_instr = mfma.selected_mfma;
......
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