Skip to content

GitLab

Commit 2963dd96 authored by aska-0096's avatar aska-0096
Browse files

temp save

parent ccb94cea
Hide whitespace changes
Inline Side-by-side
Showing with 1593 additions and 0 deletions
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View file @ 2963dd96
add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp) add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp 0 → 100644
View file @ 2963dd96
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#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/impl/device_batched_contraction_multiple_d_wmma_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/utility/numeric.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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DDataType = F16;
using DsDataType = ck::Tuple<DDataType>;
using EDataType = F16;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 1;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Add;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
using DeviceOpInstanceKKNN =
ck::tensor_operation::device::DeviceBatchedContractionMultipleD_Wmma_CShuffle<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
ck::Tuple<DDataType>,
EDataType,
AccDataType,
CShuffleDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
ABSpec,
ABSpec,
DESpec,
256,
128,
256,
8,
8,
16,
16,
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,
1,
1,
S<1, 32, 1, 8>,
8>;
using DeviceOpInstance = DeviceOpInstanceKKNN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <ck::index_t NumDimG,
ck::index_t NumDimM,
ck::index_t NumDimN,
ck::index_t NumDimK,
typename ADataType,
typename BDataType,
typename EDataType,
typename AccDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
false>
struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_gs_ms_ks,
const Tensor<BDataType>& b_gs_ns_ks,
Tensor<EDataType>& e_gs_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_gs_ms_ks_{a_gs_ms_ks},
b_gs_ns_ks_{b_gs_ns_ks},
e_gs_ms_ns_{e_gs_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_gs_ms_ks_;
const Tensor<BDataType>& b_gs_ns_ks_;
Tensor<EDataType>& e_gs_ms_ns_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
};
// Invoker
struct Invoker : public ck::tensor_operation::device::BaseInvoker
{
using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a,
ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
arg.b_element_op_(
v_b,
ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
v_acc += v_a * v_b;
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
std::thread::hardware_concurrency());
return 0;
}
float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
const StreamConfig& /* stream_config */ = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
{
return true;
}
static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
const Tensor<BDataType>& b_gs_ns_ks,
Tensor<EDataType>& e_gs_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{
a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceContraction_G2_M2_N2_K1"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::index_t G0 = 1;
ck::index_t G1 = 1;
ck::index_t M0 = 1;
ck::index_t M1 = 1;
ck::index_t N0 = 1;
ck::index_t N1 = 1;
ck::index_t K0 = 1;
// A[G0, G1, M0, M1, K0]
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
// B[G0, G1, N0, N1, K0]
std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
// D[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
// E[G0, G1, M0, N0, M1, N1]
std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
std::vector<ck::index_t> e_gs_ms_ns_strides{
G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
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);
}
std::cout<<"CP -4 "<<std::endl;
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::cout<<"CP -3 "<<std::endl;
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
break;
}
std::cout<<"CP -2 "<<std::endl;
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) *
e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
std::cout<<"CP -1 "<<std::endl;
// set zero
e_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// device operation
std::cout<<"CP 0 "<<std::endl;
auto op = DeviceOpInstance{};
auto invoker = op.MakeInvoker();
auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b_gs_ns_ks_lengths,
b_gs_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
e_gs_ms_ns_lengths,
e_gs_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
std::cout<<"CP 1 "<<std::endl;
if(!op.IsSupportedArgument(argument))
{
std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::cout<<"CP 2 "<<std::endl;
ck::index_t G =
ck::accumulate_n<ck::index_t>(e_gs_ms_ns_lengths.begin(), NumDimG, 1, std::multiplies<>{});
ck::index_t M = ck::accumulate_n<ck::index_t>(
e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * G * M * N * K;
std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< op.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
{
for(size_t g1 = 0; g1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++g1)
{
for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m0)
{
for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++m1)
{
for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n0)
{
for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
++n1)
{
cde_element_op(e_gs_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
c_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
d_gs_ms_ns(g0, g1, m0, m1, n0, n1));
}
}
}
}
}
}
return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
}
return 0;
}
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp 0 → 100644
View file @ 2963dd96
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.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/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// Tensor Contraction:
// input : A
// input : B
// input : D0, D1, ...
// output : E
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
// B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
// D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
// E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
// 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
// stride_i = stride_{i - 1} * extent_{i - 1}
// So tensor
// [G0, G1, G2, M, N]
// transposed into tensor
// [G0, G2, G1, M, N]
// with strides
// [G2 * G1 * M * N, G1 * M * N, M * N, N, 1]
// is again a packed tensor. MakeGridDescriptor() currently just merges dimensions and ignores some
// strides from input tensor extents so finer dimension information is lost. Merging dimensions is
// 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 in a traditional sense of "packed" tensor
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
GemmSpecialization GemmSpec,
TensorSpecialization ASpec,
TensorSpecialization BSpec,
TensorSpecialization DESpec,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t K0PerBlock,
ck::index_t K1,
ck::index_t MPerWMMA,
ck::index_t NPerWMMA,
ck::index_t MRepeat,
ck::index_t NRepeat,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
bool ABlockLdsAddExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BBlockLdsAddExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
ck::index_t NumPrefetch = 1,
ck::LoopScheduler LoopSched = make_default_loop_scheduler(),
ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
: public DeviceBatchedContractionMultipleD<NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
{
using DeviceOp = DeviceBatchedContractionMultipleD_Wmma_CShuffle;
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
// K1 = Max Vector Access Pixels
static constexpr auto K1Number = Number<K1>{};
static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
// Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
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)
{
assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
const auto to_tuple = [&](auto& vec, auto start, auto end) {
return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
};
const auto a_ms_ks_lengths = to_tuple(
a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
const auto a_ms_ks_strides = to_tuple(
a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
// dimension Ids for M0, M1, ...
constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
// dimension Ids for K0, K1, ...
constexpr auto kDimIds =
typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
// lengths for M0, M1, ...
const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
// lengths for K0, K1, ...
const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
if constexpr(ASpec == TensorSpecialization::Packed)
{
auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
make_tuple(M, K),
make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
}
else
{
// naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
const auto a_grid_desc_ms_ks =
make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
// transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
a_grid_desc_ms_ks,
make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
make_tuple(mDimIds, kDimIds),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
}
}
// Assume: B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
const std::vector<index_t>& b_gs_ns_ks_strides_vec)
{
assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
const auto to_tuple = [&](auto& vec, auto start, auto end) {
return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
};
const auto b_ns_ks_lengths = to_tuple(
b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
const auto b_ns_ks_strides = to_tuple(
b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
// dimension Ids for N0, N1, ...
constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
// dimension Ids for K0, K1, ...
constexpr auto kDimIds =
typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
// lengths for K0, K1, ...
const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
// lengths for N0, N1, ...
const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
if constexpr(BSpec == TensorSpecialization::Packed)
{
auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
make_tuple(N, K),
make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
}
else
{
// naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
const auto b_grid_desc_ns_ks =
make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
// transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
b_grid_desc_ns_ks,
make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
make_tuple(nDimIds, kDimIds),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
}
}
// assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
const std::vector<index_t>& e_gs_ms_ns_strides_vec)
{
assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
const auto to_tuple = [&](auto& vec, auto start, auto end) {
return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
};
const auto e_ms_ns_lengths = to_tuple(
e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
const auto e_ms_ns_strides = to_tuple(
e_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
// dimension Ids for M0, M1, ...
constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
// dimension Ids for N0, N1, ...
constexpr auto nDimIds =
typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
// lengths for M0, M1, ...
const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
// lengths for K0, K1, ...
const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
if constexpr(DESpec == TensorSpecialization::Packed)
{
auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
make_tuple(M, N),
make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
else
{
// naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
const auto e_grid_desc_ms_ns =
make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
// transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
e_grid_desc_ms_ns,
make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
make_tuple(mDimIds, nDimIds),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
}
}
// assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
const std::vector<index_t>& e_gs_ms_ns_strides_vec)
{
assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
const auto to_tuple = [&](auto& vec, auto start, auto end) {
return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
};
const auto e_gs_ms_ns_lengths =
to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
const auto e_gs_ms_ns_strides =
to_tuple(e_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(e_gs_ms_ns_lengths, gDimIds);
// lengths for M0, M1, ...
const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
// lengths for K0, K1, ...
const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
if constexpr(DESpec == 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 e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
make_tuple(G, M, N),
make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
// return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
return e_grid_desc_g_mraw_nraw;
}
else
{
// naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
const auto e_grid_desc_gs_ms_ns =
make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
// transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
// N2 * ...]
const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
e_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>{}));
// return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
return e_grid_desc_g_mraw_nraw;
}
}
static auto MakeDsGridDescriptor_M_N(
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
{
return generate_tuple(
[&](auto i) {
return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
ds_gs_ms_ns_strides_vec[i]);
},
Number<NumDTensor>{});
}
static auto MakeDsGridDescriptor_G_M_N(
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
{
return generate_tuple(
[&](auto i) {
return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
ds_gs_ms_ns_strides_vec[i]);
},
Number<NumDTensor>{});
}
// Gridwise descriptor, mapping to whole given provblem.
using AGridDesc_M_K = decltype(MakeAGridDescriptor_M_K({}, {}));
using BGridDesc_N_K = decltype(MakeBGridDescriptor_N_K({}, {}));
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N({}, {}));
using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
using EGridDesc_G_M_N = decltype(MakeEGridDescriptor_G_M_N({}, {}));
// A desc for source in blockwise copy
template <typename AGridDesc_M_K>
__host__ __device__ static constexpr auto
MakeAGridDescriptor_K0_M_K1(const AGridDesc_M_K& a_grid_desc_m_k)
{
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
const auto AK0 = K / K1;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
// B desc for source in blockwise copy
template <typename BGridDesc_N_K>
__host__ __device__ static constexpr auto
MakeBGridDescriptor_K0_N_K1(const BGridDesc_N_K& b_grid_desc_n_k)
{
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
const auto BK0 = K / K1;
return transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
struct ComputePtrOffsetOfStridedBatch
{
ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
index_t batch_stride_B,
DsGridDesc_G_M_N ds_grid_desc_g_m_n,
EGridDesc_G_M_N e_grid_desc_g_m_n)
: batch_stride_A_(batch_stride_A),
batch_stride_B_(batch_stride_B),
ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
{
}
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return static_cast<long_index_t>(g_idx) * batch_stride_A_;
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return static_cast<long_index_t>(g_idx) * batch_stride_B_;
}
__host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
{
std::array<long_index_t, NumDTensor> ds_offset;
static_for<0, NumDTensor, 1>{}([&](auto i) {
ds_offset[i] = static_cast<long_index_t>(g_idx) *
ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
});
return ds_offset;
}
__host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
{
return static_cast<long_index_t>(g_idx) *
e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
}
private:
index_t batch_stride_A_;
index_t batch_stride_B_;
DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
EGridDesc_G_M_N e_grid_desc_g_m_n_;
};
using AGridDesc_K0_M_K1 = decltype(DeviceOp::MakeAGridDescriptor_K0_M_K1(AGridDesc_M_K{}));
using BGridDesc_K0_N_K1 = decltype(DeviceOp::MakeBGridDescriptor_K0_N_K1(BGridDesc_N_K{}));
// GridwiseOp
using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
// DataType Family
ADataType,
BDataType,
AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
// InMemory Data Descriptor
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
DsGridDesc_M_N,
EGridDesc_M_N,
// ElementwiseOp Family
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
InMemoryDataOperationEnum::Set,
// Tiling Family
MPerBlock,
NPerBlock,
K0PerBlock,
MPerWMMA,
NPerWMMA,
K1,
MRepeat,
NRepeat,
// ThreadCluster Family
BlockSize,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CDEShuffleBlockTransferScalarPerVector_NPerBlock,
NumPrefetch,
LoopSched,
PipelineVer>;
// Argument
struct Argument : public BaseArgument
{
Argument(const void* p_a_grid,
const void* p_b_grid,
std::array<const void*, NumDTensor> p_ds_grid,
void* p_e_grid,
const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& b_gs_ns_ks_lengths,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
const std::vector<index_t>& e_gs_ms_ns_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_strides,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
const std::vector<index_t>& e_gs_ms_ns_strides,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
p_ds_grid_{},
p_e_grid_{static_cast<EDataType*>(p_e_grid)},
a_grid_desc_m_k_{},
b_grid_desc_n_k_{},
ds_grid_desc_m_n_{},
e_grid_desc_m_n_{},
ds_grid_desc_g_m_n_{
DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
e_grid_desc_g_m_n_{
DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
ds_grid_desc_mblock_mperblock_nblock_nperblock{},
e_grid_desc_mblock_mperblock_nblock_nperblock{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op},
a_mz_stride_{},
a_kz_stride_{},
b_nz_stride_{},
b_kz_stride_{},
ds_nz_stride_{},
e_nz_stride_{},
a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
compute_ptr_offset_of_batch_{
a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_}
{
a_grid_desc_m_k_ =
DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
a_grid_desc_m_k_ =
DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(a_grid_desc_m_k_);
b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(b_grid_desc_n_k_);
static_for<0, NumDTensor, 1>{}([&](auto i) {
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
// D pointer
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
// D desc
ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths[i],
ds_gs_ms_ns_strides[i]);
});
e_grid_desc_m_n_ =
DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
block_2_ctile_map_ = GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01);
if(GridwiseOp::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
ds_grid_desc_m_n_,
e_grid_desc_m_n_,
block_2_ctile_map_))
{
ds_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
ds_grid_desc_m_n_);
e_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
e_grid_desc_m_n_);
}
// for sanity check of vector memory access
a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
for(index_t i = 0; i < NumDTensor; ++i)
{
ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
}
e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
}
// Pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
typename GridwiseOp::DsGridPointer p_ds_grid_;
EDataType* p_e_grid_;
// Tensor Descriptors
AGridDesc_M_K a_grid_desc_m_k_;
BGridDesc_N_K b_grid_desc_n_k_;
DsGridDesc_M_N ds_grid_desc_m_n_;
EGridDesc_M_N e_grid_desc_m_n_;
DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
EGridDesc_G_M_N e_grid_desc_g_m_n_;
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock;
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock;
// Block to Tile mapping
typename GridwiseOp::DefaultBlock2CTileMap block_2_ctile_map_;
// Idle
index_t M01_;
index_t N01_;
// ElementwiseOp
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CDEElementwiseOperation cde_element_op_;
// Strides for the last M/N/K dimensions of A/B/Ds/E
// for sanity check of vector load/store
index_t a_mz_stride_;
index_t a_kz_stride_;
index_t b_nz_stride_;
index_t b_kz_stride_;
std::array<index_t, NumDTensor> ds_nz_stride_;
index_t e_mz_stride_;
index_t e_nz_stride_;
index_t a_batch_stride_;
index_t b_batch_stride_;
// Batch Offset
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
#if 0
{
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0)
<< ", " << arg.c_grid_desc_m_n_.GetLength(I1) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I2) << "}" << std::endl;
}
#endif
if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
throw std::runtime_error(
"wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
}
const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(GridwiseOp::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_contraction_multiple_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
typename GridwiseOp::DsGridPointer,
EDataType,
remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
ComputePtrOffsetOfStridedBatch,
remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
true>; // Last Option is W/O
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_ds_grid_,
arg.p_e_grid_,
G,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.compute_ptr_offset_of_batch_,
arg.block_2_ctile_map_);
}
else
{
const auto kernel = kernel_contraction_multiple_d_wmma_cshuffle<
GridwiseOp,
ADataType,
BDataType,
typename GridwiseOp::DsGridPointer,
EDataType,
remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<
typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
ComputePtrOffsetOfStridedBatch,
remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
false>;
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_ds_grid_,
arg.p_e_grid_,
G,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
arg.a_element_op_,
arg.b_element_op_,
arg.cde_element_op_,
arg.compute_ptr_offset_of_batch_,
arg.block_2_ctile_map_);
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx1100")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
return false;
}
}
else
{
return false;
}
if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.ds_grid_desc_m_n_,
arg.e_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
return false;
}
// check vector access
static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
(BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
"wrong!");
// vector memory access of A: could be on M or AK1 dimension
if constexpr(ABlockTransferSrcVectorDim == 1)
{
if(!(arg.a_mz_stride_ == 1 &&
arg.a_grid_desc_k0_m_k1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
{
return false;
}
}
else
{
if(!(arg.a_kz_stride_ == 1 &&
arg.a_grid_desc_k0_m_k1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
{
return false;
}
}
// vector memory access of B: could be on N or BK1 dimension
if constexpr(BBlockTransferSrcVectorDim == 1)
{
if(!(arg.b_nz_stride_ == 1 &&
arg.b_grid_desc_k0_n_k1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
{
return false;
}
}
else
{
if(!(arg.b_kz_stride_ == 1 &&
arg.b_grid_desc_k0_n_k1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
{
return false;
}
}
// vector memory access of Ds: always on NPerBlock dimension
bool valid_d_access = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
if(!(arg.ds_nz_stride_[i] == 1 &&
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetLength(I3) %
CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
0))
{
valid_d_access = false;
}
});
if(valid_d_access == false)
{
return false;
}
// vector memory access of E: always on NPerBlock dimension
if(!((arg.e_nz_stride_ == 1 &&
arg.e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I3) %
CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
0) ||
CDEShuffleBlockTransferScalarPerVector_NPerBlock == 1))
{
return false;
}
return true;
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto
MakeArgument(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
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::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
const std::vector<index_t>& e_gs_ms_ns_lengths,
const std::vector<index_t>& e_gs_ms_ns_strides,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{p_a,
p_b,
p_ds,
p_e,
a_gs_ms_ks_lengths,
b_gs_ns_ks_lengths,
ds_gs_ms_ns_lengths,
e_gs_ms_ns_lengths,
a_gs_ms_ks_strides,
b_gs_ns_ks_strides,
ds_gs_ms_ns_strides,
e_gs_ms_ns_strides,
1,
1,
a_element_op,
b_element_op,
cde_element_op};
}
// polymorphic
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
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::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
const std::vector<index_t>& e_gs_ms_ns_lengths,
const std::vector<index_t>& e_gs_ms_ns_strides,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) override
{
return std::make_unique<Argument>(p_a,
p_b,
p_ds,
p_e,
a_gs_ms_ks_lengths,
b_gs_ns_ks_lengths,
ds_gs_ms_ns_lengths,
e_gs_ms_ns_lengths,
a_gs_ms_ks_strides,
b_gs_ns_ks_strides,
ds_gs_ms_ns_strides,
e_gs_ms_ns_strides,
1,
1,
a_element_op,
b_element_op,
cde_element_op);
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
std::map<LoopScheduler, std::string> LoopSchedToString{
{LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
{PipelineVersion::v2, "v2"}};
// clang-format off
str << "DeviceBatchedContractionMultipleD_Wmma_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock << ", "
<< K1 << ", "
<< MPerWMMA << ", "
<< NPerWMMA << ", "
<< MRepeat << ", "
<< NRepeat
<< ">"
<< " NumPrefetch: "
<< NumPrefetch << ", "
<< "LoopScheduler: "
<< LoopSchedToString[LoopSched] << ", "
<< "PipelineVersion: "
<< PipelineVersionToString[PipelineVer];
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
View file @ 2963dd96
...@@ -17,6 +17,98 @@ ...@@ -17,6 +17,98 @@
namespace ck { namespace ck {
template <typename GridwiseOp,
typename ADataType,
typename BDataType,
typename DsPointer,
typename EDataType,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputePtrOffsetOfBatch,
typename Block2CTileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_contraction_multiple_d_wmma_cshuffle(
const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const index_t batch_count,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
e_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2CTileMap block_2_etile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
DsPointer p_ds_grid_grp;
static constexpr index_t NumDTensor =
DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
static_for<0, NumDTensor, 1>{}(
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_ds_grid_grp,
p_e_grid + e_batch_offset,
p_shared,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
cde_element_op,
block_2_etile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_ds_grid;
ignore = p_e_grid;
ignore = batch_count;
ignore = a_element_op;
ignore = b_element_op;
ignore = cde_element_op;
ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = block_2_etile_map;
ignore = compute_ptr_offset_of_batch;
#endif
}
template <typename GridwiseOp, template <typename GridwiseOp,
typename ADataType, typename ADataType,
typename BDataType, typename BDataType,
......
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