Skip to content

GitLab

Commit 7e7640ce authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into cpu_avx2

parents cc8df39e f015c776
Hide whitespace changes
Inline Side-by-side
Showing with 831 additions and 14 deletions
library/src/obselete_driver_offline/conv_fwd_driver_offline_nchwc.cpp
View file @ 7e7640ce
...@@ -282,7 +282,7 @@ int main(int argc, char* argv[]) ...@@ -282,7 +282,7 @@ int main(int argc, char* argv[])
print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w)); print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w));
print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w)); print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w));
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
......
library/src/obselete_driver_offline/conv_maxpool_fwd_driver_offline_nchwc.cpp
View file @ 7e7640ce
...@@ -300,7 +300,7 @@ int main(int argc, char* argv[]) ...@@ -300,7 +300,7 @@ int main(int argc, char* argv[])
print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w)); print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w));
print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w)); print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w));
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
......
library/src/obselete_driver_offline/conv_wrw_driver_offline.cpp
View file @ 7e7640ce
...@@ -289,7 +289,7 @@ int main(int argc, char* argv[]) ...@@ -289,7 +289,7 @@ int main(int argc, char* argv[])
print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w)); print_array("ConvStrides", make_tuple(conv_stride_h, conv_stride_w));
print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w)); print_array("ConvDilations", make_tuple(conv_dilation_h, conv_dilation_w));
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
......
library/src/obselete_driver_offline/gemm_driver_offline.cpp
View file @ 7e7640ce
...@@ -313,7 +313,7 @@ int main(int argc, char* argv[]) ...@@ -313,7 +313,7 @@ int main(int argc, char* argv[])
ostream_HostTensorDescriptor(b.mDesc, std::cout << "b: "); ostream_HostTensorDescriptor(b.mDesc, std::cout << "b: ");
ostream_HostTensorDescriptor(c_host.mDesc, std::cout << "c: "); ostream_HostTensorDescriptor(c_host.mDesc, std::cout << "c: ");
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 7e7640ce
...@@ -39,3 +39,4 @@ add_subdirectory(conv2d_bwd_data) ...@@ -39,3 +39,4 @@ add_subdirectory(conv2d_bwd_data)
add_subdirectory(reduce) add_subdirectory(reduce)
add_subdirectory(convnd_bwd_data) add_subdirectory(convnd_bwd_data)
add_subdirectory(grouped_gemm) add_subdirectory(grouped_gemm)
add_subdirectory(batched_gemm_reduce)
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt 0 → 100644
View file @ 7e7640ce
set(DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
)
add_instance_library(device_batched_gemm_reduce_instance ${DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE})
install(TARGETS device_batched_gemm_reduce_instance LIBRARY DESTINATION lib)
clang_tidy_check(device_batched_gemm_reduce_instance)
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp 0 → 100644
View file @ 7e7640ce
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp 0 → 100644
View file @ 7e7640ce
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 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, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 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, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 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, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 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, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp 0 → 100644
View file @ 7e7640ce
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp 0 → 100644
View file @ 7e7640ce
#include <stdlib.h>
#include "config.hpp"
#include "device_batched_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, 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, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ 7e7640ce
...@@ -35,6 +35,7 @@ set(PROFILER_SOURCE ...@@ -35,6 +35,7 @@ set(PROFILER_SOURCE
src/profile_convnd_bwd_data.cpp src/profile_convnd_bwd_data.cpp
src/profile_reduce.cpp src/profile_reduce.cpp
src/profile_grouped_gemm.cpp src/profile_grouped_gemm.cpp
src/profile_batched_gemm_reduce.cpp
) )
add_executable(ckProfiler ${PROFILER_SOURCE}) add_executable(ckProfiler ${PROFILER_SOURCE})
...@@ -54,3 +55,4 @@ target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance) ...@@ -54,3 +55,4 @@ target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
profiler/include/profile_batched_gemm_impl.hpp
View file @ 7e7640ce
#pragma once #pragma once
#include <memory> #include <memory>
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "tensor_layout.hpp"
#include "device.hpp"
#include "host_tensor_generator.hpp"
#include "device_gemm.hpp"
#include "reference_batched_gemm.hpp" #include "reference_batched_gemm.hpp"
namespace ck { namespace ck {
...@@ -51,7 +59,7 @@ template <typename ADataType, ...@@ -51,7 +59,7 @@ template <typename ADataType,
typename ALayout, typename ALayout,
typename BLayout, typename BLayout,
typename CLayout> typename CLayout>
void profile_batched_gemm_impl(int do_verification, bool profile_batched_gemm_impl(int do_verification,
int init_method, int init_method,
bool do_log, bool do_log,
int nrepeat, int nrepeat,
...@@ -63,6 +71,8 @@ void profile_batched_gemm_impl(int do_verification, ...@@ -63,6 +71,8 @@ void profile_batched_gemm_impl(int do_verification,
int StrideC, int StrideC,
int BatchCount = 1) int BatchCount = 1)
{ {
bool pass = true;
auto f_host_tensor_descriptor = [](std::size_t batch_count, auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row, std::size_t row,
std::size_t col, std::size_t col,
...@@ -93,7 +103,7 @@ void profile_batched_gemm_impl(int do_verification, ...@@ -93,7 +103,7 @@ void profile_batched_gemm_impl(int do_verification,
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl; std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl; std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
...@@ -378,12 +388,14 @@ void profile_batched_gemm_impl(int do_verification, ...@@ -378,12 +388,14 @@ void profile_batched_gemm_impl(int do_verification,
{ {
bf16_to_f32_(c_g_m_n_device_result, *c_f32_g_m_n_device_result); bf16_to_f32_(c_g_m_n_device_result, *c_f32_g_m_n_device_result);
check_error(*c_f32_g_m_n_host_result, *c_f32_g_m_n_device_result); float err = check_error(*c_f32_g_m_n_host_result, *c_f32_g_m_n_device_result);
pass = pass && (err < 1E-6);
} }
else else
{ {
check_error(c_g_m_n_host_result, c_g_m_n_device_result); float err = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
pass = pass && (err < 1E-6);
} }
if(do_log) if(do_log)
...@@ -407,6 +419,8 @@ void profile_batched_gemm_impl(int do_verification, ...@@ -407,6 +419,8 @@ void profile_batched_gemm_impl(int do_verification,
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, " std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl; << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
return pass;
} }
} // namespace profiler } // namespace profiler
......
profiler/include/profile_batched_gemm_reduce_impl.hpp 0 → 100644
View file @ 7e7640ce
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "device_gemm_reduce.hpp"
#include "reference_batched_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::ReduceSum,
ck::tensor_operation::element_wise::ReduceSquareSum>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_batched_gemm_reduce_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int BatchCount)
{
bool pass = true;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({row * stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({col * stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
std::cout << "d0_g_m: " << d0_g_m_host_result.mDesc << std::endl;
std::cout << "d1_g_m: " << d1_g_m_host_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency();
switch(init_method)
{
case 0: break;
case 1:
std::srand(0);
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
std::srand(0);
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::tensor_operation::element_wise::ReduceSum;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
if(do_verification)
{
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
for(int batch = 0; batch < BatchCount; ++batch)
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReduceZeroValue();
float d1_acc = d1_reduce_op.GetReduceZeroValue();
for(int n = 0; n < N; ++n)
{
d0_reduce_op.Reduce(d0_acc, c_g_m_n_host_result(batch, m, n));
d1_reduce_op.Reduce(d1_acc, c_g_m_n_host_result(batch, m, n));
}
d0_g_m_host_result(batch, m) = d0_acc;
d1_g_m_host_result(batch, m) = d1_acc;
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmReduceNoOpPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
d0_reduce_op,
d1_reduce_op,
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// warm up
invoker_ptr->Run(argument_ptr.get());
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker_ptr->Run(argument_ptr.get());
timer.End();
total_time += timer.GetElapsedTime();
}
float ave_time = total_time / nrepeat;
std::string gemm_name = gemm_ptr->GetTypeString();
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
sizeof(BDataType) * BatchCount * K * N +
sizeof(CDataType) * BatchCount * 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, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
float c_error = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
float d0_error = check_error(d0_g_m_host_result, d0_g_m_device_result);
float d1_error = check_error(d1_g_m_host_result, d1_g_m_device_result);
pass = pass && (c_error < 1E-6);
pass = pass && (d0_error < 1E-6);
pass = pass && (d1_error < 1E-6);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_host: ", d0_g_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "d0_device: ", d0_g_m_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_host: ", d1_g_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "d1_device: ", d1_g_m_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_bias_2d_impl.hpp
View file @ 7e7640ce
...@@ -98,7 +98,7 @@ void profile_gemm_bias_2d_impl(int do_verification, ...@@ -98,7 +98,7 @@ void profile_gemm_bias_2d_impl(int do_verification,
std::cout << "c0_m_n: " << c0_m_n.mDesc << std::endl; std::cout << "c0_m_n: " << c0_m_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl; std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
......
profiler/include/profile_gemm_bias_relu_impl.hpp
View file @ 7e7640ce
...@@ -83,7 +83,7 @@ void profile_gemm_bias_relu_impl(int do_verification, ...@@ -83,7 +83,7 @@ void profile_gemm_bias_relu_impl(int do_verification,
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl; std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl; std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
......
profiler/include/profile_gemm_impl.hpp
View file @ 7e7640ce
...@@ -120,7 +120,7 @@ void profile_gemm_impl(int do_verification, ...@@ -120,7 +120,7 @@ void profile_gemm_impl(int do_verification,
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl; std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
...@@ -408,6 +408,10 @@ void profile_gemm_impl(int do_verification, ...@@ -408,6 +408,10 @@ void profile_gemm_impl(int do_verification,
if(gemm_ptr->IsSupportedArgument(argument_ptr.get())) if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{ {
// re-init C to zero before profiling next kernel
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
std::string gemm_name = gemm_ptr->GetTypeString(); std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat); float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ 7e7640ce
...@@ -98,7 +98,7 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -98,7 +98,7 @@ bool profile_gemm_reduce_impl(int do_verification,
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl; std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl; std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
......
profiler/include/profile_grouped_gemm_impl.hpp
View file @ 7e7640ce
...@@ -95,7 +95,7 @@ void profile_grouped_gemm_impl(int do_verification, ...@@ -95,7 +95,7 @@ void profile_grouped_gemm_impl(int do_verification,
<< "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl; << "]:" << c_m_n_device_results[i].mDesc << std::endl;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
......
profiler/include/profile_reduce_impl.hpp
View file @ 7e7640ce
...@@ -242,7 +242,7 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -242,7 +242,7 @@ void profile_reduce_impl_impl(bool do_verification,
size_t invariant_total_length = out.mDesc.GetElementSize(); size_t invariant_total_length = out.mDesc.GetElementSize();
size_t reduce_total_length = in.mDesc.GetElementSize() / invariant_total_length; size_t reduce_total_length = in.mDesc.GetElementSize() / invariant_total_length;
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = 1;
if(do_verification) if(do_verification)
{ {
......
profiler/src/profile_batched_gemm_reduce.cpp 0 → 100644
View file @ 7e7640ce
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_batched_gemm_reduce_impl.hpp"
int profile_batched_gemm_reduce(int argc, char* argv[])
{
enum struct GemmMatrixLayout_t
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmReduceDataType_t
{
F32_F32_F32_F32_F32, // 0
F16_F16_F16_F32_F32, // 1
};
if(!(argc == 15 || argc == 16))
{
printf("arg1: tensor operation (batched_gemm: BatchedGEMM+Reduce)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg7: run kernel # of times (>1)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
printf("arg15: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmReduceDataType_t>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout_t>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const int nrepeat = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int BatchCount = std::stoi(argv[14]);
if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::MK_KN_MN)
{
ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::MK_NK_MN)
{
ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::KM_KN_MN)
{
ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout_t::KM_NK_MN)
{
ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
nrepeat,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 1;
}
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