Skip to content

GitLab

Commit e57bd240 authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into add_fp16_wmma_conv_instance
parents 4b70d68e 37a8c1f7
Hide whitespace changes
Inline Side-by-side
Showing with 632 additions and 126 deletions
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
View file @ e57bd240
...@@ -16,7 +16,7 @@ namespace ck { ...@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_BF16
void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances( void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>& DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
...@@ -36,7 +36,8 @@ void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances( ...@@ -36,7 +36,8 @@ void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>& DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances( void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
...@@ -56,7 +57,8 @@ void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances( ...@@ -56,7 +57,8 @@ void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances( void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
...@@ -76,7 +78,8 @@ void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances( ...@@ -76,7 +78,8 @@ void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances( void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<DeviceBatchedGemm<Col, std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
Row, Row,
...@@ -120,7 +123,7 @@ void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances( ...@@ -120,7 +123,7 @@ void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
template <typename ALayout, template <typename ALayout,
typename BLayout, typename BLayout,
typename CLayout, typename CLayout,
...@@ -151,7 +154,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -151,7 +154,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> && if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<CDataType, float>) is_same_v<CDataType, float>)
{ {
...@@ -176,8 +179,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -176,8 +179,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(op_ptrs); add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> && #endif
is_same_v<CDataType, half_t>) #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{ {
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> && if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
...@@ -200,8 +205,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -200,8 +205,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs); add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> && #endif
is_same_v<CDataType, bhalf_t>) #ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
is_same_v<CDataType, bhalf_t>)
{ {
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> && if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
...@@ -224,8 +231,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -224,8 +231,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(op_ptrs); add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> && #endif
is_same_v<CDataType, int8_t>) #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
is_same_v<CDataType, int8_t>)
{ {
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> && if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
...@@ -248,7 +257,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -248,7 +257,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(op_ptrs); add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
View file @ e57bd240
...@@ -14,7 +14,7 @@ ...@@ -14,7 +14,7 @@
using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu; using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
using CDE1ElementOp = ck::tensor_operation::element_wise::Add; using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
#ifdef CK_ENABLE_FP16
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -137,3 +137,4 @@ struct DeviceOperationInstanceFactory< ...@@ -137,3 +137,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
View file @ e57bd240
...@@ -13,7 +13,7 @@ ...@@ -13,7 +13,7 @@
#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/device_operation_instance_factory.hpp" #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef CK_ENABLE_FP16
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -91,3 +91,4 @@ struct DeviceOperationInstanceFactory< ...@@ -91,3 +91,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
View file @ e57bd240
...@@ -16,7 +16,7 @@ namespace ck { ...@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP16
void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances( void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2, DeviceBatchedGemmSoftmaxGemmPermute<2,
...@@ -58,7 +58,8 @@ void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_ ...@@ -58,7 +58,8 @@ void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_
PassThrough, PassThrough,
MaskingSpecialization::MaskDisabled>>>& MaskingSpecialization::MaskDisabled>>>&
instances); instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances( void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2, DeviceBatchedGemmSoftmaxGemmPermute<2,
...@@ -100,7 +101,7 @@ void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf ...@@ -100,7 +101,7 @@ void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf
PassThrough, PassThrough,
MaskingSpecialization::MaskDisabled>>>& MaskingSpecialization::MaskDisabled>>>&
instances); instances);
#endif
template <typename ADataType, template <typename ADataType,
typename B0DataType, typename B0DataType,
typename B1DataType, typename B1DataType,
...@@ -147,7 +148,7 @@ struct DeviceOperationInstanceFactory< ...@@ -147,7 +148,7 @@ struct DeviceOperationInstanceFactory<
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> && if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> && is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> &&
Acc0BiasDataType::Size() == 1 && Acc0BiasDataType::Size() == 1 &&
...@@ -164,6 +165,8 @@ struct DeviceOperationInstanceFactory< ...@@ -164,6 +165,8 @@ struct DeviceOperationInstanceFactory<
op_ptrs); op_ptrs);
} }
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> && else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> && is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
Acc0BiasDataType::Size() == 1 && Acc0BiasDataType::Size() == 1 &&
...@@ -180,6 +183,7 @@ struct DeviceOperationInstanceFactory< ...@@ -180,6 +183,7 @@ struct DeviceOperationInstanceFactory<
op_ptrs); op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
View file @ e57bd240
...@@ -16,7 +16,7 @@ namespace ck { ...@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP16
void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance( void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row, std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
Col, Col,
...@@ -111,3 +111,4 @@ struct DeviceOperationInstanceFactory< ...@@ -111,3 +111,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
View file @ e57bd240
...@@ -14,11 +14,12 @@ ...@@ -14,11 +14,12 @@
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef DL_KERNELS
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP16
void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances( void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col, std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
Row, Row,
...@@ -122,7 +123,8 @@ void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instan ...@@ -122,7 +123,8 @@ void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instan
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances( void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col, std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
Row, Row,
...@@ -226,7 +228,7 @@ void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances ...@@ -226,7 +228,7 @@ void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
template <typename ALayout, template <typename ALayout,
typename BLayout, typename BLayout,
typename ELayout, typename ELayout,
...@@ -261,7 +263,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -261,7 +263,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> && if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>) is_same_v<EDataType, half_t>)
{ {
...@@ -294,6 +296,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -294,6 +296,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
op_ptrs); op_ptrs);
} }
} }
#endif
#ifdef CK_ENABLE_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<EDataType, int8_t>) is_same_v<EDataType, int8_t>)
{ {
...@@ -326,7 +330,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -326,7 +330,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
op_ptrs); op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
...@@ -335,3 +339,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche ...@@ -335,3 +339,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
View file @ e57bd240
...@@ -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/device_operation_instance_factory.hpp" #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef CK_ENABLE_FP16
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -119,3 +119,4 @@ struct DeviceOperationInstanceFactory< ...@@ -119,3 +119,4 @@ struct DeviceOperationInstanceFactory<
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
View file @ e57bd240
...@@ -16,7 +16,7 @@ namespace ck { ...@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP16
void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances( void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2, DeviceBatchedGemmSoftmaxGemmPermute<2,
...@@ -58,7 +58,8 @@ void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_g ...@@ -58,7 +58,8 @@ void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_g
PassThrough, PassThrough,
MaskingSpecialization::MaskDisabled>>>& MaskingSpecialization::MaskDisabled>>>&
instances); instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances( void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2, DeviceBatchedGemmSoftmaxGemmPermute<2,
...@@ -100,6 +101,7 @@ void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf ...@@ -100,6 +101,7 @@ void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf
PassThrough, PassThrough,
MaskingSpecialization::MaskDisabled>>>& MaskingSpecialization::MaskDisabled>>>&
instances); instances);
#endif
template <typename ADataType, template <typename ADataType,
typename B0DataType, typename B0DataType,
...@@ -146,7 +148,7 @@ struct DeviceOperationInstanceFactory< ...@@ -146,7 +148,7 @@ struct DeviceOperationInstanceFactory<
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> && if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>) is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
{ {
...@@ -161,6 +163,8 @@ struct DeviceOperationInstanceFactory< ...@@ -161,6 +163,8 @@ struct DeviceOperationInstanceFactory<
op_ptrs); op_ptrs);
} }
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> && else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16>) is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16>)
{ {
...@@ -175,6 +179,7 @@ struct DeviceOperationInstanceFactory< ...@@ -175,6 +179,7 @@ struct DeviceOperationInstanceFactory<
op_ptrs); op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
View file @ e57bd240
...@@ -16,7 +16,7 @@ namespace ck { ...@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP32
// float // float
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -65,7 +65,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn ...@@ -65,7 +65,8 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear>>>& instances);
#endif
#ifdef CK_ENABLE_FP64
// double // double
void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance( void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -114,7 +115,7 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn ...@@ -114,7 +115,7 @@ void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn
PassThrough, PassThrough,
PassThrough, PassThrough,
Bilinear>>>& instances); Bilinear>>>& instances);
#endif
// Contraction + Bilinear // Contraction + Bilinear
template <index_t NumDimM, template <index_t NumDimM,
index_t NumDimN, index_t NumDimN,
...@@ -149,7 +150,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -149,7 +150,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> && if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<DDataType, float> && is_same_v<EDataType, float>) is_same_v<DDataType, float> && is_same_v<EDataType, float>)
{ {
...@@ -165,7 +166,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -165,7 +166,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
op_ptrs); op_ptrs);
} }
} }
#endif
#ifdef CK_ENABLE_FP64
if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> && if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
is_same_v<DDataType, double> && is_same_v<EDataType, double>) is_same_v<DDataType, double> && is_same_v<EDataType, double>)
{ {
...@@ -181,7 +183,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -181,7 +183,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
op_ptrs); op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
View file @ e57bd240
...@@ -16,7 +16,7 @@ namespace ck { ...@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP32
// float // float
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -65,7 +65,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instanc ...@@ -65,7 +65,8 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instanc
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale>>>& instances);
#endif
#ifdef CK_ENABLE_FP64
// double // double
void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance( void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
std::vector<std::unique_ptr<DeviceContractionMultipleD<2, std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
...@@ -114,7 +115,7 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instanc ...@@ -114,7 +115,7 @@ void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instanc
PassThrough, PassThrough,
PassThrough, PassThrough,
Scale>>>& instances); Scale>>>& instances);
#endif
// Contraction + Scale // Contraction + Scale
template <index_t NumDimM, template <index_t NumDimM,
index_t NumDimN, index_t NumDimN,
...@@ -148,7 +149,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -148,7 +149,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
static auto GetInstances() static auto GetInstances()
{ {
std::vector<std::unique_ptr<DeviceOp>> op_ptrs; std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> && if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<EDataType, float>) is_same_v<EDataType, float>)
{ {
...@@ -164,7 +165,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -164,7 +165,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
op_ptrs); op_ptrs);
} }
} }
#endif
#ifdef CK_ENABLE_FP64
if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> && if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
is_same_v<EDataType, double>) is_same_v<EDataType, double>)
{ {
...@@ -180,7 +182,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra ...@@ -180,7 +182,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContra
op_ptrs); op_ptrs);
} }
} }
#endif
return op_ptrs; return op_ptrs;
} }
}; };
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
View file @ e57bd240
...@@ -16,7 +16,7 @@ namespace ck { ...@@ -16,7 +16,7 @@ namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_BF16
// conv1d backward data // conv1d backward data
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances( void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<1, std::vector<std::unique_ptr<DeviceConvBwdData<1,
...@@ -29,17 +29,20 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances( ...@@ -29,17 +29,20 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances( void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceConvBwdData<1, NWC, KXC, NWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceConvBwdData<1, NWC, KXC, NWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances( 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__ #endif
#ifdef CK_ENABLE_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,
...@@ -52,6 +55,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances( ...@@ -52,6 +55,7 @@ void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
#ifdef CK_ENABLE_BF16
// 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,
...@@ -64,7 +68,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances( ...@@ -64,7 +68,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances( void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<2, std::vector<std::unique_ptr<DeviceConvBwdData<2,
NHWC, NHWC,
...@@ -76,7 +81,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances( ...@@ -76,7 +81,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances( void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<2, std::vector<std::unique_ptr<DeviceConvBwdData<2,
NHWC, NHWC,
...@@ -88,7 +94,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances( ...@@ -88,7 +94,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#ifdef __int8__ #endif
#ifdef CK_ENABLE_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,
...@@ -101,6 +108,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances( ...@@ -101,6 +108,8 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
#ifdef DL_KERNELS
#ifdef CK_ENABLE_FP16
// 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,
...@@ -113,7 +122,8 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances( ...@@ -113,7 +122,8 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances( void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<2, std::vector<std::unique_ptr<DeviceConvBwdData<2,
NHWC, NHWC,
...@@ -125,7 +135,8 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances( ...@@ -125,7 +135,8 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#ifdef __int8__ #endif
#ifdef CK_ENABLE_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,
...@@ -138,6 +149,8 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances( ...@@ -138,6 +149,8 @@ void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
#endif
#ifdef CK_ENABLE_BF16
// 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,
...@@ -150,7 +163,8 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances( ...@@ -150,7 +163,8 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances( void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<3, std::vector<std::unique_ptr<DeviceConvBwdData<3,
NDHWC, NDHWC,
...@@ -162,7 +176,8 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances( ...@@ -162,7 +176,8 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances( void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
std::vector<std::unique_ptr<DeviceConvBwdData<3, std::vector<std::unique_ptr<DeviceConvBwdData<3,
NDHWC, NDHWC,
...@@ -174,7 +189,8 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances( ...@@ -174,7 +189,8 @@ void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#ifdef __int8__ #endif
#ifdef CK_ENABLE_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,
...@@ -229,20 +245,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -229,20 +245,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
{ {
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs); add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #ifdef CK_ENABLE_FP16
is_same_v<OutDataType, half_t>) if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(op_ptrs); add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> && #ifdef CK_ENABLE_BF16
is_same_v<OutDataType, ck::bhalf_t>) if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{ {
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__ #endif
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && #ifdef CK_ENABLE_INT8
is_same_v<OutDataType, int8_t>) if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, 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);
} }
...@@ -255,26 +274,35 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -255,26 +274,35 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs); add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs); add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
#endif
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #ifdef CK_ENABLE_FP16
is_same_v<OutDataType, half_t>) if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs); add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(op_ptrs); add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
#endif
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> && #ifdef CK_ENABLE_BF16
is_same_v<OutDataType, ck::bhalf_t>) if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{ {
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__ #endif
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && #ifdef CK_ENABLE_INT8
is_same_v<OutDataType, int8_t>) if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, 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);
#ifdef DL_KERNELS
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
} }
#endif #endif
} }
...@@ -286,20 +314,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw ...@@ -286,20 +314,23 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBw
{ {
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs); add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #ifdef CK_ENABLE_FP16
is_same_v<OutDataType, half_t>) if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(op_ptrs); add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> && #ifdef CK_ENABLE_BF16
is_same_v<OutDataType, ck::bhalf_t>) if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{ {
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__ #endif
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && #ifdef CK_ENABLE_INT8
is_same_v<OutDataType, int8_t>) if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, 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);
} }
......
library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
View file @ e57bd240
...@@ -18,11 +18,17 @@ namespace device { ...@@ -18,11 +18,17 @@ namespace device {
namespace instance { namespace instance {
// conv2d forward // conv2d forward
#ifdef CK_ENABLE_FP16
void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances( void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<std::unique_ptr<
DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances( void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceConvFwd<2, std::vector<std::unique_ptr<DeviceConvFwd<2,
NHWC, NHWC,
...@@ -34,17 +40,14 @@ void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances( ...@@ -34,17 +40,14 @@ void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances( #ifdef CK_ENABLE_FP32
std::vector<std::unique_ptr<
DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances( void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceConvFwd<2, NHWC, KYXC, NHWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceConvFwd<2, NHWC, KYXC, NHWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances( void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<std::unique_ptr<DeviceConvFwd<2, std::vector<std::unique_ptr<DeviceConvFwd<2,
NHWC, NHWC,
...@@ -56,6 +59,7 @@ void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances( ...@@ -56,6 +59,7 @@ void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_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,
...@@ -99,23 +103,29 @@ struct DeviceOperationInstanceFactory< ...@@ -99,23 +103,29 @@ struct DeviceOperationInstanceFactory<
{ {
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
} }
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs); add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_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_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs); add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
} }
#endif
} }
return op_ptrs; return op_ptrs;
......
library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
View file @ e57bd240
...@@ -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/device_operation_instance_factory.hpp" #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef CK_ENABLE_FP16
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -77,3 +77,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElemen ...@@ -77,3 +77,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElemen
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
View file @ e57bd240
...@@ -17,11 +17,17 @@ namespace tensor_operation { ...@@ -17,11 +17,17 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#if defined(CK_ENABLE_FP16) && defined(DL_KERNELS)
void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances( void add_device_gemm_dl_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>>>&
instances); instances);
void add_device_gemm_dpp_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances( void add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_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>>>&
...@@ -32,6 +38,11 @@ void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances( ...@@ -32,6 +38,11 @@ void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dpp_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances( void add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
...@@ -42,6 +53,11 @@ void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances( ...@@ -42,6 +53,11 @@ void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances( void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
...@@ -52,15 +68,20 @@ void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances( ...@@ -52,15 +68,20 @@ void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances( void add_device_gemm_dpp_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>>>&
instances); instances);
void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(
std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#if defined(CK_ENABLE_FP32) && defined(DL_KERNELS)
void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances( void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances( void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(
...@@ -77,7 +98,8 @@ void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances( ...@@ -77,7 +98,8 @@ 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__ #endif
#if defined(CK_ENABLE_INT8) && defined(DL_KERNELS)
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>>>&
...@@ -117,7 +139,8 @@ void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances( ...@@ -117,7 +139,8 @@ void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
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);
#endif
#ifdef CK_ENABLE_INT8
void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_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>>>&
...@@ -138,90 +161,93 @@ void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances( ...@@ -138,90 +161,93 @@ void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_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);
#endif #endif
#ifdef CK_ENABLE_FP16
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>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_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>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances( void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances( void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_xdl_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>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances( #endif
#ifdef CK_ENABLE_BF16
void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_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, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif
void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances( #ifdef CK_ENABLE_FP32
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_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, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_f32_f32_f32_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, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances( void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(
...@@ -243,7 +269,8 @@ void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances( ...@@ -243,7 +269,8 @@ void add_device_gemm_xdl_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);
#endif
#ifdef CK_ENABLE_FP64
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances( void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
...@@ -264,7 +291,7 @@ void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances( ...@@ -264,7 +291,7 @@ void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(
std::vector<std::unique_ptr< std::vector<std::unique_ptr<
DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>& DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
#endif
template <typename ALayout, template <typename ALayout,
typename BLayout, typename BLayout,
typename CLayout, typename CLayout,
...@@ -303,31 +330,40 @@ struct DeviceOperationInstanceFactory< ...@@ -303,31 +330,40 @@ struct DeviceOperationInstanceFactory<
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
} }
} }
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> && else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>) is_same_v<CDataType, half_t>)
{ {
...@@ -335,16 +371,22 @@ struct DeviceOperationInstanceFactory< ...@@ -335,16 +371,22 @@ struct DeviceOperationInstanceFactory<
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
} }
...@@ -352,19 +394,27 @@ struct DeviceOperationInstanceFactory< ...@@ -352,19 +394,27 @@ struct DeviceOperationInstanceFactory<
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_km_kn_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);
add_device_gemm_dpp_f16_f16_f16_km_nk_mn_instances(op_ptrs);
#endif
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
} }
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> && else if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
is_same_v<CDataType, ck::bhalf_t>) is_same_v<CDataType, ck::bhalf_t>)
{ {
...@@ -389,7 +439,8 @@ struct DeviceOperationInstanceFactory< ...@@ -389,7 +439,8 @@ 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__ #endif
#ifdef CK_ENABLE_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>)
{ {
...@@ -397,29 +448,37 @@ struct DeviceOperationInstanceFactory< ...@@ -397,29 +448,37 @@ struct DeviceOperationInstanceFactory<
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
add_device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances(op_ptrs);
#endif
} }
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(op_ptrs);
#endif
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(op_ptrs);
add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances(op_ptrs);
#endif
} }
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> && else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>) is_same_v<CLayout, Row>)
{ {
add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs);
#ifdef DL_KERNELS
add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(op_ptrs);
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
} }
} }
#endif #endif
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
View file @ e57bd240
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef CK_ENABLE_FP16
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -170,3 +170,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu ...@@ -170,3 +170,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
View file @ e57bd240
...@@ -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/device_operation_instance_factory.hpp" #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#ifdef CK_ENABLE_FP16
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -220,3 +220,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu ...@@ -220,3 +220,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
View file @ e57bd240
...@@ -10,7 +10,7 @@ ...@@ -10,7 +10,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp" #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#ifdef CK_ENABLE_FP16
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -136,3 +136,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu ...@@ -136,3 +136,4 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp 0 → 100644
View file @ e57bd240
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
MultiplyAdd>>>&);
void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Col,
Row_Row_Tuple,
Row,
F16,
F16,
F16_F16_Tuple,
F16,
PassThrough,
PassThrough,
MultiplyAdd>>>&);
void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Row_Row_Tuple,
Row,
F16,
F8,
F32_F32_Tuple,
F16,
PassThrough,
PassThrough,
MultiplyAdd>>>&);
void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Col,
Row_Row_Tuple,
Row,
F16,
F8,
F32_F32_Tuple,
F16,
PassThrough,
PassThrough,
MultiplyAdd>>>&);
// GEMM + Multiply + Add
template <typename ALayout,
typename BLayout,
typename D0Layout,
typename D1Layout,
typename ELayout,
typename ADataType,
typename BDataType,
typename D0DataType,
typename D1DataType,
typename EDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::MultiplyAdd>>
{
using DeviceOp = DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<D0Layout, D1Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType, D1DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::MultiplyAdd>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
op_ptrs);
}
}
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<D0DataType, float> && is_same_v<D1DataType, float> &&
is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
View file @ e57bd240
...@@ -57,6 +57,46 @@ void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances( ...@@ -57,6 +57,46 @@ void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
DeviceGemmSplitK<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>& DeviceGemmSplitK<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
instances); instances);
void add_device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
template <typename ADataType, template <typename ADataType,
typename BDataType, typename BDataType,
typename CDataType, typename CDataType,
...@@ -136,6 +176,54 @@ struct DeviceOperationInstanceFactory< ...@@ -136,6 +176,54 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
} }
} }
else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs);
}
}
return op_ptrs; return op_ptrs;
} }
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp 0 → 100644
View file @ e57bd240
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_FP16
void add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmStreamK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmStreamK<
ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
{
using DeviceOp = DeviceGemmStreamK<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#if 0
if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
is_same_v<CDataType, float>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
}
}
#endif
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // 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