Skip to content

GitLab

Commit f0fd0263 authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'amd-develop' into amd-master

parents 4e911f3e a8fafc3f
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2012 additions and 739 deletions
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp
View file @ f0fd0263
...@@ -195,17 +195,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -195,17 +195,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false> template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N( static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
ck::index_t N, const ck::index_t N,
ck::index_t K, const ck::index_t K,
ck::index_t C, const ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides, const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations, const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads, const std::array<ck::index_t, NDimSpatial>& input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads, const std::array<ck::index_t, NDimSpatial>& input_right_pads,
ck::index_t batch_k) const ck::index_t batch_k)
{ {
using namespace ck; using namespace ck;
...@@ -347,17 +347,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -347,17 +347,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
} // function end } // function end
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false> template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N( static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
ck::index_t N, const ck::index_t N,
ck::index_t K, const ck::index_t K,
ck::index_t C, const ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides, const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations, const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads, const std::array<ck::index_t, NDimSpatial>& input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads, const std::array<ck::index_t, NDimSpatial>& input_right_pads,
ck::index_t batch_k) const ck::index_t batch_k)
{ {
using namespace ck; using namespace ck;
...@@ -515,17 +515,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -515,17 +515,17 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false> template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N( static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
ck::index_t N, const ck::index_t N,
ck::index_t K, const ck::index_t K,
ck::index_t C, const ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides, const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations, const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
std::array<ck::index_t, NDimSpatial> input_left_pads, const std::array<ck::index_t, NDimSpatial>& input_left_pads,
std::array<ck::index_t, NDimSpatial> input_right_pads, const std::array<ck::index_t, NDimSpatial>& input_right_pads,
ck::index_t batch_k) const ck::index_t batch_k)
{ {
using namespace ck; using namespace ck;
...@@ -784,17 +784,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -784,17 +784,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
Argument(const InDataType* p_in_grid, Argument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid, WeiDataType* p_wei_grid,
const OutDataType* p_out_grid, const OutDataType* p_out_grid,
ck::index_t G, const ck::index_t G,
ck::index_t N, const ck::index_t N,
ck::index_t K, const ck::index_t K,
ck::index_t C, const ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides, const std::array<ck::index_t, NDimSpatial + 3>& /*input_strides*/,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations, const std::array<ck::index_t, NDimSpatial + 3>& /*output_strides*/,
std::array<ck::index_t, NDimSpatial> input_left_pads, const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
std::array<ck::index_t, NDimSpatial> input_right_pads, const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
const std::array<ck::index_t, NDimSpatial>& input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
...@@ -897,18 +899,18 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -897,18 +899,18 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
InElementwiseOperation c_element_op_; InElementwiseOperation c_element_op_;
// for checking IsSupportedArgument() // for checking IsSupportedArgument()
index_t Conv_G_; const index_t Conv_G_;
index_t Conv_N_; const index_t Conv_N_;
index_t Conv_K_; const index_t Conv_K_;
index_t Conv_C_; const index_t Conv_C_;
std::array<ck::index_t, NDimSpatial> input_spatial_lengths_; const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_; const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> output_spatial_lengths_; const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths_;
std::array<ck::index_t, NDimSpatial> conv_filter_strides_; const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
std::array<ck::index_t, NDimSpatial> conv_filter_dilations_; const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations_;
std::array<ck::index_t, NDimSpatial> input_left_pads_; const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
std::array<ck::index_t, NDimSpatial> input_right_pads_; const std::array<ck::index_t, NDimSpatial>& input_right_pads_;
index_t k_batch_; index_t k_batch_;
}; };
...@@ -1111,17 +1113,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -1111,17 +1113,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
static auto MakeArgument(const InDataType* p_in_grid, static auto MakeArgument(const InDataType* p_in_grid,
WeiDataType* p_wei_grid, WeiDataType* p_wei_grid,
const OutDataType* p_out_grid, const OutDataType* p_out_grid,
ck::index_t G, const ck::index_t G,
ck::index_t N, const ck::index_t N,
ck::index_t K, const ck::index_t K,
ck::index_t C, const ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides, const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations, const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
std::array<ck::index_t, NDimSpatial> input_left_pads, const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
std::array<ck::index_t, NDimSpatial> input_right_pads, const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
const std::array<ck::index_t, NDimSpatial>& input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
...@@ -1137,6 +1141,8 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -1137,6 +1141,8 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
input_spatial_lengths, input_spatial_lengths,
filter_spatial_lengths, filter_spatial_lengths,
output_spatial_lengths, output_spatial_lengths,
input_strides,
output_strides,
conv_filter_strides, conv_filter_strides,
conv_filter_dilations, conv_filter_dilations,
input_left_pads, input_left_pads,
...@@ -1153,17 +1159,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -1153,17 +1159,19 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
MakeArgumentPointer(const void* p_in_grid, MakeArgumentPointer(const void* p_in_grid,
void* p_wei_grid, void* p_wei_grid,
const void* p_out_grid, const void* p_out_grid,
ck::index_t G, const ck::index_t G,
ck::index_t N, const ck::index_t N,
ck::index_t K, const ck::index_t K,
ck::index_t C, const ck::index_t C,
std::array<ck::index_t, NDimSpatial> input_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
std::array<ck::index_t, NDimSpatial> output_spatial_lengths, const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
std::array<ck::index_t, NDimSpatial> conv_filter_strides, const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
std::array<ck::index_t, NDimSpatial> conv_filter_dilations, const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
std::array<ck::index_t, NDimSpatial> input_left_pads, const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
std::array<ck::index_t, NDimSpatial> input_right_pads, const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
const std::array<ck::index_t, NDimSpatial>& input_left_pads,
const std::array<ck::index_t, NDimSpatial>& input_right_pads,
InElementwiseOperation in_element_op, InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op, WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op, OutElementwiseOperation out_element_op,
...@@ -1179,6 +1187,8 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl ...@@ -1179,6 +1187,8 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
input_spatial_lengths, input_spatial_lengths,
filter_spatial_lengths, filter_spatial_lengths,
output_spatial_lengths, output_spatial_lengths,
input_strides,
output_strides,
conv_filter_strides, conv_filter_strides,
conv_filter_dilations, conv_filter_dilations,
input_left_pads, input_left_pads,
......
include/ck/utility/type_convert.hpp
View file @ f0fd0263
...@@ -62,24 +62,6 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, half_t>(half_ ...@@ -62,24 +62,6 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, half_t>(half_
return type_convert<bhalf_t>(x_fp32); return type_convert<bhalf_t>(x_fp32);
} }
// convert bfp16 to int32 via fp32
template <>
inline __host__ __device__ constexpr int32_t type_convert<int32_t, bhalf_t>(bhalf_t x)
{
float x_fp32 = type_convert<float>(x);
return static_cast<int32_t>(x_fp32);
}
// convert int32 to bfp16 via fp32
template <>
inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int32_t>(int32_t x)
{
float x_fp32 = static_cast<float>(x);
return type_convert<bhalf_t>(x_fp32);
}
// convert bfp16 to int8 via fp32 // convert bfp16 to int8 via fp32
template <> template <>
inline __host__ __device__ constexpr int8_t type_convert<int8_t, bhalf_t>(bhalf_t x) inline __host__ __device__ constexpr int8_t type_convert<int8_t, bhalf_t>(bhalf_t x)
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
View file @ f0fd0263
...@@ -39,7 +39,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances( ...@@ -39,7 +39,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#ifdef __int8__
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances( void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<1, std::vector<std::unique_ptr<DeviceConvBwdData<1,
NWC, NWC,
...@@ -51,7 +51,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances( ...@@ -51,7 +51,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// conv2d backward data // conv2d backward data
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances( void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<2, std::vector<std::unique_ptr<DeviceConvBwdData<2,
...@@ -88,7 +88,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances( ...@@ -88,7 +88,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#ifdef __int8__
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances( void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<2, std::vector<std::unique_ptr<DeviceConvBwdData<2,
NHWC, NHWC,
...@@ -100,7 +100,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances( ...@@ -100,7 +100,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// conv2d dl // conv2d dl
void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances( void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<2, std::vector<std::unique_ptr<DeviceConvBwdData<2,
...@@ -125,7 +125,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances( ...@@ -125,7 +125,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#ifdef __int8__
void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances( void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<2, std::vector<std::unique_ptr<DeviceConvBwdData<2,
NHWC, NHWC,
...@@ -137,6 +137,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances( ...@@ -137,6 +137,7 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// conv3d backward data // conv3d backward data
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances( void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<3, std::vector<std::unique_ptr<DeviceConvBwdData<3,
...@@ -173,7 +174,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances( ...@@ -173,7 +174,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#ifdef __int8__
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances( void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<3, std::vector<std::unique_ptr<DeviceConvBwdData<3,
NDHWC, NDHWC,
...@@ -185,7 +186,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances( ...@@ -185,7 +186,7 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
typename InLayout, typename InLayout,
typename WeiLayout, typename WeiLayout,
...@@ -239,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -239,11 +240,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
{ {
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(op_ptrs); add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(op_ptrs);
} }
#ifdef __int8__
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t>)
{ {
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(op_ptrs); add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(op_ptrs);
} }
#endif
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> && else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>) is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
...@@ -266,12 +269,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -266,12 +269,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
{ {
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs); add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
} }
#ifdef __int8__
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t>)
{ {
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs); add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(op_ptrs); add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
} }
#endif
} }
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> && else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>) is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
...@@ -292,11 +297,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -292,11 +297,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
{ {
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(op_ptrs); add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(op_ptrs);
} }
#ifdef __int8__
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>) is_same_v<OutDataType, int8_t>)
{ {
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(op_ptrs); add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(op_ptrs);
} }
#endif
} }
return op_ptrs; return op_ptrs;
......
library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
View file @ f0fd0263
...@@ -77,7 +77,7 @@ void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances( ...@@ -77,7 +77,7 @@ void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#ifdef __int8__
void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances( void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
...@@ -118,6 +118,27 @@ void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances( ...@@ -118,6 +118,27 @@ void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(
DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
...@@ -183,26 +204,6 @@ void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances( ...@@ -183,26 +204,6 @@ void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances( void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
...@@ -388,6 +389,7 @@ struct DeviceOperationInstanceFactory< ...@@ -388,6 +389,7 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
} }
} }
#ifdef __int8__
else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> && else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
is_same_v<CDataType, int8_t>) is_same_v<CDataType, int8_t>)
{ {
...@@ -420,7 +422,7 @@ struct DeviceOperationInstanceFactory< ...@@ -420,7 +422,7 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
View file @ f0fd0263
...@@ -16,7 +16,7 @@ namespace device { ...@@ -16,7 +16,7 @@ namespace device {
namespace instance { namespace instance {
// conv2d backward data // conv2d backward data
void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances( void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK, GNHWK,
GKYXC, GKYXC,
...@@ -30,7 +30,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances( ...@@ -30,7 +30,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances( void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK, GNHWK,
GKYXC, GKYXC,
...@@ -44,7 +44,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances( ...@@ -44,7 +44,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances( void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK, GNHWK,
GKYXC, GKYXC,
...@@ -58,7 +58,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances( ...@@ -58,7 +58,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances( void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK, NHWGK,
GKYXC, GKYXC,
...@@ -72,7 +72,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances( ...@@ -72,7 +72,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances( void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK, NHWGK,
GKYXC, GKYXC,
...@@ -86,7 +86,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances( ...@@ -86,7 +86,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK, NHWGK,
GKYXC, GKYXC,
...@@ -100,6 +100,91 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( ...@@ -100,6 +100,91 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
// conv3d backward data
void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
typename OutLayout, typename OutLayout,
typename WeiLayout, typename WeiLayout,
...@@ -139,43 +224,96 @@ struct DeviceOperationInstanceFactory< ...@@ -139,43 +224,96 @@ struct DeviceOperationInstanceFactory<
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 2)
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
{ {
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
{ is_same_v<OutLayout, GNHWK>)
add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>)
{
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>)
{
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
op_ptrs);
}
} }
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutDataType, BF16>) is_same_v<OutLayout, NHWGK>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs); if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
op_ptrs);
}
} }
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> && else if constexpr(NumDimSpatial == 3)
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
{ {
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
{ is_same_v<OutLayout, GNDHWK>)
add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs); if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>)
{
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>)
{
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
op_ptrs);
}
} }
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutDataType, BF16>) is_same_v<OutLayout, NDHWGK>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs); if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>)
{
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>)
{
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
op_ptrs);
}
} }
} }
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View file @ f0fd0263
...@@ -91,6 +91,42 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances( ...@@ -91,6 +91,42 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
BF16,
F32,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// conv3d backward weight // conv3d backward weight
void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
...@@ -128,6 +164,42 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances ...@@ -128,6 +164,42 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
BF16,
F32,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F32,
F32,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
typename InLayout, typename InLayout,
typename WeiLayout, typename WeiLayout,
...@@ -162,66 +234,126 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -162,66 +234,126 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> && if constexpr(NumDimSpatial == 1)
is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>)
{ {
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InLayout, GNWC> && is_same_v<WeiLayout, GKXC> &&
is_same_v<OutDataType, float>) is_same_v<OutLayout, GNWK>)
{ {
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs); if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
} is_same_v<OutDataType, float>)
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && {
is_same_v<OutDataType, half_t>) add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
{ }
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs); else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
} is_same_v<OutDataType, half_t>)
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && {
is_same_v<OutDataType, ck::bhalf_t>) add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
{ }
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
op_ptrs); is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
op_ptrs);
}
} }
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> && else if constexpr(NumDimSpatial == 2)
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
{ {
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutDataType, float>) is_same_v<OutLayout, GNHWK>)
{ {
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
op_ptrs);
}
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutDataType, half_t>) is_same_v<OutLayout, NHWGK>)
{ {
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
} is_same_v<OutDataType, float>)
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && {
is_same_v<OutDataType, ck::bhalf_t>) add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
{ op_ptrs);
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( }
op_ptrs); else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
op_ptrs);
}
} }
} }
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> && else if constexpr(NumDimSpatial == 3)
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>)
{ {
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutDataType, float>) is_same_v<OutLayout, GNDHWK>)
{
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances( if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
op_ptrs); is_same_v<OutDataType, float>)
{
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
op_ptrs);
}
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && is_same_v<WeiDataType, float> && else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutLayout, NDHWGK>)
{ {
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
op_ptrs); is_same_v<OutDataType, float>)
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
}
else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
op_ptrs);
}
} }
} }
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ f0fd0263
...@@ -12,9 +12,42 @@ set(CK_DEVICE_INSTANCES) ...@@ -12,9 +12,42 @@ set(CK_DEVICE_INSTANCES)
FOREACH(subdir_path ${dir_list}) FOREACH(subdir_path ${dir_list})
set(target_dir) set(target_dir)
IF(IS_DIRECTORY "${subdir_path}") IF(IS_DIRECTORY "${subdir_path}")
get_filename_component(target_dir ${subdir_path} NAME) set(cmake_instance)
add_subdirectory(${target_dir}) file(READ "${subdir_path}/CMakeLists.txt" cmake_instance)
list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>) set(add_inst 0)
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8")
#message("fp8 instance found!")
set(add_inst 1)
endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16")
#message("fp16 instance found!")
set(add_inst 1)
endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32")
#message("fp32 instance found!")
set(add_inst 1)
endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64")
#message("fp64 instance found!")
set(add_inst 1)
endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16")
#message("bf16 instance found!")
set(add_inst 1)
endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8")
#message("int8 instance found!")
set(add_inst 1)
endif()
if(NOT "${cmake_instance}" MATCHES "DTYPES")
#message("instance should be built for all types!")
set(add_inst 1)
endif()
if(add_inst EQUAL 1 OR NOT DEFINED DTYPES)
get_filename_component(target_dir ${subdir_path} NAME)
add_subdirectory(${target_dir})
list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
endif()
ENDIF() ENDIF()
ENDFOREACH() ENDFOREACH()
......
library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
View file @ f0fd0263
add_instance_library(device_batched_gemm_multi_d_instance set(BATCHED_GEMM_MULTID_INSTANCES)
device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp)
device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp)
device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp)
device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp)
device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp endif()
device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp)
device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp)
) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp)
endif()
add_instance_library(device_batched_gemm_multi_d_instance ${BATCHED_GEMM_MULTID_INSTANCES})
library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
View file @ f0fd0263
add_instance_library(device_conv2d_bwd_data_instance set(CONV2D_BWD_DATA_INSTANCES)
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp)
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp)
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp)
device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp endif()
device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
) list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp)
list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp)
list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp)
endif()
add_instance_library(device_conv2d_bwd_data_instance ${CONV2D_BWD_DATA_INSTANCES})
library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
View file @ f0fd0263
...@@ -11,7 +11,7 @@ ...@@ -11,7 +11,7 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#ifdef __int8__
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -151,3 +151,4 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances( ...@@ -151,3 +151,4 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
View file @ f0fd0263
add_instance_library(device_gemm_instance set(GEMM_INSTANCES)
device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp)
device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp)
device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp)
device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp)
device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp endif()
device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp)
device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp)
device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp)
device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp)
device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp)
device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp endif()
device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp)
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp)
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp)
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp)
device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp)
device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp)
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp)
device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp)
device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp)
device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp)
device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp)
device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp)
device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp)
device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp)
device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp)
device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp)
list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp)
endif()
device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp add_instance_library(device_gemm_instance ${GEMM_INSTANCES})
device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
)
set(ENABLE_PIPELINE_V2_OPT OFF) set(ENABLE_PIPELINE_V2_OPT OFF)
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
View file @ f0fd0263
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#ifdef __int8__
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -80,3 +80,4 @@ void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances( ...@@ -80,3 +80,4 @@ void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
View file @ f0fd0263
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#ifdef __int8__
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -78,3 +78,4 @@ void add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances( ...@@ -78,3 +78,4 @@ void add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances(
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
View file @ f0fd0263
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#ifdef __int8__
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -80,3 +80,4 @@ void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances( ...@@ -80,3 +80,4 @@ void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
View file @ f0fd0263
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#ifdef __int8__
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -78,3 +78,4 @@ void add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances( ...@@ -78,3 +78,4 @@ void add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
View file @ f0fd0263
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#ifdef __int8__
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -80,3 +80,4 @@ void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances( ...@@ -80,3 +80,4 @@ void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
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