Skip to content

GitLab

"git@developer.sourcefind.cn:gaoqiong/composable_kernel.git" did not exist on "ad5db169b615b6c7015a670848de29c4a14a71d3"
Unverified Commit 72b7ae25 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Update staging branch. (#706) 



* update daily build from rocm 5.4.3 to 5.5 (#693)

* Fix grouped_gemm_splitk kernels on MI300. (#694)

* replace amd_buffer_atomic_add with hip_atomic_add

* fix grouped_gemm_splitk kernels on mi300

* fix syntax

* revert experimental atomic_add changes

---------
Co-authored-by: default avatarJing Zhang <jizhan@amd.com>

* Fix the group of quantization_int8 kernels on MI300. (#695)

* replace amd_buffer_atomic_add with hip_atomic_add

* fix grouped_gemm_splitk kernels on mi300

* fix syntax

* revert experimental atomic_add changes

* fix the group of kernels from ticket 723 on MI300

---------
Co-authored-by: default avatarJing Zhang <jizhan@amd.com>

* Optimize bf16 conversion (#664)

* Add TypeConvert class and start refactoring

* Refactor TypeConvert as a struct

* Get back to template functions type_convert

* Add a type_convert_bf16_rtn, set rtz as default

* Clean up

* Add UnaryConvertPrecision struct for high-precision workloads

* Format

* Update type_convert to UnaryConvert on threadwise level

* Update UnaryConvertPrecision

* Format

* Fix chmod

* Add a flag to pick converion method

* Format

* Remove the added flag

* Merge elementwise op with type conversion

* Move type_convert to elemwise op, update the op

* Update type_convert_precision -> bf16_convert_rtn

* Clean up

* Update comments

* Update the CK_WORKAROUND_DENORM_FIX flag handling

* Update the unneeded op to work but warn user

* Remove the message

* Use a PassThrough instead of ConvertBF16RTN to calcaulate reference

* Format

* Add missing include

* Normalization/split k (#615)

* Add contraction profiler and tests (#701)

* Add contraction profiler and tests

* Build and style fixes

* Allow to use any elementwise operator for ref_contraction

* Introduce profile_contraction_scale and profile_contraction_bilinear

* Make ref_contraction generic and extend interface tests

* Stylistic minor fixes

* Extend test_contraction_interface

---------
Co-authored-by: default avatarJing Zhang <jizhan@amd.com>
Co-authored-by: default avatarRostyslav Geyyer <46627076+geyyer@users.noreply.github.com>
Co-authored-by: default avatarrocking <ChunYu.Lai@amd.com>
Co-authored-by: default avatarBartłomiej Kocot <bartlomiejkocot98@gmail.com>
parent bbe74503
Hide whitespace changes
Inline Side-by-side
Showing with 245 additions and 668 deletions
Jenkinsfile
View file @ 72b7ae25
...@@ -19,7 +19,7 @@ def runShell(String command){ ...@@ -19,7 +19,7 @@ def runShell(String command){
def getDockerImageName(){ def getDockerImageName(){
def img def img
if (params.ROCMVERSION != "5.5" && params.ROCMVERSION != "5.6"){ if (params.ROCMVERSION != "5.6"){
if (params.COMPILER_VERSION == "") { if (params.COMPILER_VERSION == "") {
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}" img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
} }
...@@ -597,7 +597,7 @@ def process_results(Map conf=[:]){ ...@@ -597,7 +597,7 @@ def process_results(Map conf=[:]){
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true
0 21 * * * % ROCMVERSION=5.4.3;COMPILER_VERSION=release;COMPILER_COMMIT= 0 21 * * * % ROCMVERSION=5.5;COMPILER_VERSION=release;COMPILER_COMMIT=
0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : "" 0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
pipeline { pipeline {
......
example/15_grouped_gemm/run_grouped_gemm_example.inc
View file @ 72b7ae25
...@@ -147,7 +147,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co ...@@ -147,7 +147,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
#else #else
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data()); a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data()); b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
c_tensors_device[i]->SetZero(); c_tensors_device[i]->SetZero();
#endif #endif
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer()); p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
......
example/26_contraction/contraction_bilinear_xdl_fp32.cpp
View file @ 72b7ae25
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/numeric.hpp" #include "ck/library/utility/numeric.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -74,141 +75,6 @@ using DeviceOpInstanceMNNN = ck::tensor_operation::device:: ...@@ -74,141 +75,6 @@ using DeviceOpInstanceMNNN = ck::tensor_operation::device::
using DeviceOpInstance = DeviceOpInstanceKKNN; using DeviceOpInstance = DeviceOpInstanceKKNN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <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<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_ms_ks_{a_ms_ks},
b_ns_ks_{b_ns_ks},
e_ms_ns_{e_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_ms_ks_;
const Tensor<BDataType>& b_ns_ks_;
Tensor<EDataType>& e_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_M2_N2_K2::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
for(int k1 = 0; k1 < K1; ++k1)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
arg.b_element_op_(
v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
v_acc += v_a * v_b;
}
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_ms_ns_.mDesc.GetLengths()[0],
arg.e_ms_ns_.mDesc.GetLengths()[1],
arg.e_ms_ns_.mDesc.GetLengths()[2],
arg.e_ms_ns_.mDesc.GetLengths()[3])(
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_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{a_ms_ks, b_ns_ks, e_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_M2_N2_K2"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = true; bool do_verification = true;
...@@ -385,22 +251,22 @@ int main(int argc, char* argv[]) ...@@ -385,22 +251,22 @@ int main(int argc, char* argv[])
{ {
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides); Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM, using ReferenceOpInstance =
NumDimN, ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimK, NumDimN,
ADataType, NumDimK,
BDataType, ADataType,
CShuffleDataType, BDataType,
AccDataType, CShuffleDataType,
AElementOp, AccDataType,
BElementOp, AElementOp,
PassThrough>; BElementOp>;
auto ref_gemm = ReferenceOpInstance{}; auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument =
a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{}); ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
......
example/26_contraction/contraction_bilinear_xdl_fp64.cpp
View file @ 72b7ae25
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/numeric.hpp" #include "ck/library/utility/numeric.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -74,141 +75,6 @@ using DeviceOpInstanceMNNN = ck::tensor_operation::device:: ...@@ -74,141 +75,6 @@ using DeviceOpInstanceMNNN = ck::tensor_operation::device::
using DeviceOpInstance = DeviceOpInstanceKKNN; using DeviceOpInstance = DeviceOpInstanceKKNN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <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<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_ms_ks_{a_ms_ks},
b_ns_ks_{b_ns_ks},
e_ms_ns_{e_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_ms_ks_;
const Tensor<BDataType>& b_ns_ks_;
Tensor<EDataType>& e_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_M2_N2_K2::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
for(int k1 = 0; k1 < K1; ++k1)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
arg.b_element_op_(
v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
v_acc += v_a * v_b;
}
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_ms_ns_.mDesc.GetLengths()[0],
arg.e_ms_ns_.mDesc.GetLengths()[1],
arg.e_ms_ns_.mDesc.GetLengths()[2],
arg.e_ms_ns_.mDesc.GetLengths()[3])(
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_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{a_ms_ks, b_ns_ks, e_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_M2_N2_K2"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = true; bool do_verification = true;
...@@ -385,22 +251,22 @@ int main(int argc, char* argv[]) ...@@ -385,22 +251,22 @@ int main(int argc, char* argv[])
{ {
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides); Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM, using ReferenceOpInstance =
NumDimN, ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimK, NumDimN,
ADataType, NumDimK,
BDataType, ADataType,
CShuffleDataType, BDataType,
AccDataType, CShuffleDataType,
AElementOp, AccDataType,
BElementOp, AElementOp,
PassThrough>; BElementOp>;
auto ref_gemm = ReferenceOpInstance{}; auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument =
a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{}); ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
......
example/26_contraction/contraction_scale_xdl_fp32.cpp
View file @ 72b7ae25
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/numeric.hpp" #include "ck/library/utility/numeric.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -73,141 +74,6 @@ using DeviceOpInstanceMNN = ck::tensor_operation::device:: ...@@ -73,141 +74,6 @@ using DeviceOpInstanceMNN = ck::tensor_operation::device::
using DeviceOpInstance = DeviceOpInstanceKKN; using DeviceOpInstance = DeviceOpInstanceKKN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <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<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_ms_ks_{a_ms_ks},
b_ns_ks_{b_ns_ks},
e_ms_ns_{e_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_ms_ks_;
const Tensor<BDataType>& b_ns_ks_;
Tensor<EDataType>& e_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_M2_N2_K2::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
for(int k1 = 0; k1 < K1; ++k1)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
arg.b_element_op_(
v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
v_acc += v_a * v_b;
}
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_ms_ns_.mDesc.GetLengths()[0],
arg.e_ms_ns_.mDesc.GetLengths()[1],
arg.e_ms_ns_.mDesc.GetLengths()[2],
arg.e_ms_ns_.mDesc.GetLengths()[3])(
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_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{a_ms_ks, b_ns_ks, e_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_M2_N2_K2"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = true; bool do_verification = true;
...@@ -368,22 +234,23 @@ int main(int argc, char* argv[]) ...@@ -368,22 +234,23 @@ int main(int argc, char* argv[])
{ {
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides); Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM, using ReferenceOpInstance =
NumDimN, ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimK, NumDimN,
ADataType, NumDimK,
BDataType, ADataType,
CShuffleDataType, BDataType,
AccDataType, CShuffleDataType,
AElementOp, AccDataType,
BElementOp, AElementOp,
PassThrough>; BElementOp>;
auto ref_gemm = ReferenceOpInstance{}; auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{}); auto ref_argument =
ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
......
example/26_contraction/contraction_scale_xdl_fp64.cpp
View file @ 72b7ae25
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/numeric.hpp" #include "ck/library/utility/numeric.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -73,141 +74,6 @@ using DeviceOpInstanceMNN = ck::tensor_operation::device:: ...@@ -73,141 +74,6 @@ using DeviceOpInstanceMNN = ck::tensor_operation::device::
using DeviceOpInstance = DeviceOpInstanceKKN; using DeviceOpInstance = DeviceOpInstanceKKN;
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <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<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
{
// Argument
struct Argument : public ck::tensor_operation::device::BaseArgument
{
Argument(const Tensor<ADataType>& a_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
: a_ms_ks_{a_ms_ks},
b_ns_ks_{b_ns_ks},
e_ms_ns_{e_ms_ns},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
{
}
const Tensor<ADataType>& a_ms_ks_;
const Tensor<BDataType>& b_ns_ks_;
Tensor<EDataType>& e_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_M2_N2_K2::Argument;
float Run(const Argument& arg)
{
auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
AccDataType v_acc = 0;
for(int k0 = 0; k0 < K0; ++k0)
{
for(int k1 = 0; k1 < K1; ++k1)
{
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(
v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
arg.b_element_op_(
v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
v_acc += v_a * v_b;
}
}
AccDataType v_c;
arg.cde_element_op_(v_c, v_acc);
arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
};
make_ParallelTensorFunctor(f_ms_ns,
arg.e_ms_ns_.mDesc.GetLengths()[0],
arg.e_ms_ns_.mDesc.GetLengths()[1],
arg.e_ms_ns_.mDesc.GetLengths()[2],
arg.e_ms_ns_.mDesc.GetLengths()[3])(
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_ms_ks,
const Tensor<BDataType>& b_ns_ks,
Tensor<EDataType>& e_ms_ns,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op)
{
return Argument{a_ms_ks, b_ns_ks, e_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_M2_N2_K2"
<< std::endl;
// clang-format on
return str.str();
}
};
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
bool do_verification = true; bool do_verification = true;
...@@ -368,22 +234,23 @@ int main(int argc, char* argv[]) ...@@ -368,22 +234,23 @@ int main(int argc, char* argv[])
{ {
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides); Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM, using ReferenceOpInstance =
NumDimN, ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimK, NumDimN,
ADataType, NumDimK,
BDataType, ADataType,
CShuffleDataType, BDataType,
AccDataType, CShuffleDataType,
AElementOp, AccDataType,
BElementOp, AElementOp,
PassThrough>; BElementOp>;
auto ref_gemm = ReferenceOpInstance{}; auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{}); auto ref_argument =
ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
......
example/27_layernorm/CMakeLists.txt
View file @ 72b7ae25
add_example_executable(example_layernorm_blockwise layernorm_blockwise.cpp) add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
example/27_layernorm/common.hpp 0 → 100644
View file @ 72b7ae25
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
example/27_layernorm/layernorm_fp16.cpp 0 → 100644
View file @ 72b7ae25
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
#include "run_layernorm_example.inc"
int main() { return run_groupnorm_example<DeviceInstance>(); }
example/27_layernorm/layernorm_splitk_fp16.cpp 0 → 100644
View file @ 72b7ae25
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationSplitKImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // XScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // YScalarPerVector
#include "run_layernorm_example.inc"
int main() { return run_groupnorm_example<DeviceInstance>(); }
example/27_layernorm/layernorm_blockwise.cpp → example/27_layernorm/run_layernorm_example.inc
View file @ 72b7ae25
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #pragma once
#include <numeric>
#include <initializer_list> template <typename DeviceInstance>
#include <cstdlib> int run_groupnorm_example()
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
int main()
{ {
bool time_kernel = false; bool time_kernel = false;
...@@ -111,6 +63,10 @@ int main() ...@@ -111,6 +63,10 @@ int main()
return 1; return 1;
}; };
size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = device_instance.MakeInvokerPointer(); auto invoker_ptr = device_instance.MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel}); invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
...@@ -133,7 +89,8 @@ int main() ...@@ -133,7 +89,8 @@ int main()
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data()); y_dev.FromDevice(y.mData.data());
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results d1", 1e-3, 1e-3); pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
} }
return (pass ? 0 : 1); return (pass ? 0 : 1);
} }
example/42_groupnorm/CMakeLists.txt
View file @ 72b7ae25
add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp) add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp) add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
example/42_groupnorm/common.hpp
View file @ 72b7ae25
...@@ -12,6 +12,7 @@ ...@@ -12,6 +12,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp" #include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp" #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/fill.hpp" #include "ck/library/utility/fill.hpp"
......
example/42_groupnorm/groupnorm_splitk_fp16.cpp 0 → 100644
View file @ 72b7ae25
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationSplitKImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
YElementOp,
Rank,
NumReduceDim,
256, // BlockSize
1, // ClusterM
256, // ClusterK
1, // SliceM
16, // SliceK
1, // SrcVecDim (0=M, 1=K)
2, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
#include "run_groupnorm_example.inc"
int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
example/42_groupnorm/run_groupnorm_example.inc
View file @ 72b7ae25
...@@ -73,6 +73,10 @@ int run_groupnorm_example(int argc, char* argv[]) ...@@ -73,6 +73,10 @@ int run_groupnorm_example(int argc, char* argv[])
return 1; return 1;
}; };
size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = device_instance.MakeInvokerPointer(); auto invoker_ptr = device_instance.MakeInvokerPointer();
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, true}); float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, true});
......
include/ck/ck.hpp
View file @ 72b7ae25
...@@ -175,7 +175,10 @@ ...@@ -175,7 +175,10 @@
// denorm test fix, required to work around dissue // denorm test fix, required to work around dissue
#ifndef CK_WORKAROUND_DENORM_FIX #ifndef CK_WORKAROUND_DENORM_FIX
#define CK_WORKAROUND_DENORM_FIX 0 #define CK_WORKAROUND_DENORM_FIX 0
#endif #elif
// enable only on MI200
#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
#endif // CK_WORKAROUND_DENORM_FIX
namespace ck { namespace ck {
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ 72b7ae25
...@@ -135,7 +135,7 @@ __global__ void ...@@ -135,7 +135,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch) const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
defined(__gfx90a__) || defined(__gfx908__)) defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx940__))
// offset base pointer for each work-group // offset base pointer for each work-group
const index_t num_blocks_per_batch = const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
...@@ -710,7 +710,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK ...@@ -710,7 +710,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
// check device // check device
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" || if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908")) ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908" ||
ck::get_device_name() == "gfx940"))
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
View file @ 72b7ae25
...@@ -51,7 +51,7 @@ __global__ void ...@@ -51,7 +51,7 @@ __global__ void
const Block2CTileMap block_2_ctile_map) const Block2CTileMap block_2_ctile_map)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx1030__)) defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__))
constexpr index_t shared_block_size = constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType); GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
...@@ -552,7 +552,8 @@ struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout, ...@@ -552,7 +552,8 @@ struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" || if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" ||
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx1030") ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx940")
{ {
return GridwiseGemm::CheckValidity( return GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_); arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_);
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
View file @ 72b7ae25
...@@ -807,7 +807,7 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle ...@@ -807,7 +807,7 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
// workspace for welford intermediate mean // workspace for welford intermediate mean
workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64; workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
// workspace for welford intermediate mean // workspace for welford intermediate variance
workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64; workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
// workspace for welford intermediate count // workspace for welford intermediate count
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
View file @ 72b7ae25
...@@ -34,7 +34,8 @@ __global__ void ...@@ -34,7 +34,8 @@ __global__ void
kernel_grouped_gemm_xdl_splitk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const, kernel_grouped_gemm_xdl_splitk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
const index_t group_count) const index_t group_count)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__)) #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte(); constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
__shared__ uint8_t p_shared[shared_size]; __shared__ uint8_t p_shared[shared_size];
......
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