Skip to content

GitLab

Commit 363b6744 authored by mtgu0705's avatar mtgu0705
Browse files

add instance for gemm_ab_scale

parent 9dac9713
Hide whitespace changes
Inline Side-by-side
Showing with 237 additions and 218 deletions
library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp
View file @ 363b6744
...@@ -17,7 +17,7 @@ namespace tensor_operation { ...@@ -17,7 +17,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8)) #if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -28,14 +28,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i ...@@ -28,14 +28,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -46,14 +46,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_ ...@@ -46,14 +46,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -64,14 +64,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding ...@@ -64,14 +64,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -82,14 +82,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin ...@@ -82,14 +82,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -100,14 +100,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default ...@@ -100,14 +100,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -118,14 +118,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin ...@@ -118,14 +118,14 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -136,7 +136,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd ...@@ -136,7 +136,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -163,7 +163,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu ...@@ -163,7 +163,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
B1DataType, B1DataType,
Tuple<>, Tuple<>,
CDataType, CDataType,
128, 1,
128, 128,
128, 128,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
...@@ -180,7 +180,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu ...@@ -180,7 +180,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
B1DataType, B1DataType,
Tuple<>, Tuple<>,
CDataType, CDataType,
128, 1,
128, 128,
128, 128,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
...@@ -198,20 +198,20 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu ...@@ -198,20 +198,20 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && 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_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances( add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances(
op_ptrs); op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances( add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances(
op_ptrs); op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances( add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances(
op_ptrs); op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances( add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances(
op_ptrs); op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances( add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances(
op_ptrs); op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances( add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances(
op_ptrs); op_ptrs);
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances( add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances(
op_ptrs); op_ptrs);
} }
} }
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128.hpp
View file @ 363b6744
...@@ -34,7 +34,7 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; ...@@ -34,7 +34,7 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
template <GemmSpecialization GemmSpec> template <GemmSpecialization GemmSpec>
using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances = std::tuple< using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances = std::tuple<
// clang-format off // clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
...@@ -45,15 +45,15 @@ using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances = ...@@ -45,15 +45,15 @@ using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances =
// Spill in current compiler // Spill in current compiler
// DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, // DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 128, 16, 16, 16, 16, 7, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
// DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, // DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 224, 128, 16, 16, 16, 16, 8, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 2, 1, S<1, 64, 1, 4>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 128, 64, 128, 16, 16, 32, 32, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8> DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8, F32, F8, F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 1, 128, 128, 64, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on // clang-format on
>; >;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec> template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances = std::tuple< using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances = std::tuple<
// clang-format off // clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm| //################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| Scale| Scale| Scale| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| //################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
...@@ -61,22 +61,22 @@ using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances = s ...@@ -61,22 +61,22 @@ using device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances = s
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Latency friendly // Latency friendly
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1, F8>,
// Memory friendly // Memory friendly
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 128, 32, 128, 16, 16, 32, 32, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 128, 32, 128, 16, 16, 32, 32, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 128, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 128, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 64, 32, 128, 16, 16, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 64, 32, 128, 16, 16, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 64, 16, 128, 16, 16, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 64, 16, 128, 16, 16, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<2, 2, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 64, 16, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 64, 16, 16, 16, 16, 1, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 128, 128, 128, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 1, 128, 128, 16, 16, 128, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 32, 128, 16, 16, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 64, 128, 16, 16, 16, 16, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 64, 128, 16, 16, 16, 16, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 64, 128, 16, 16, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 16, 128, 128, 16, 16, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>, DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 16, 128, 128, 16, 16, 16, 16, 1, 4, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<4, 4, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>,
DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8> DeviceGemmMultiD_ABScale_Xdl_CShuffle_V3< Row, Col, Tuple<>, Row, F8,F32, F8,F32, Tuple<>, BF16, F32, F32, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 1, 128, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2, F8>
// clang-format on // clang-format on
>; >;
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instance.cpp
View file @ 363b6744
...@@ -8,7 +8,7 @@ namespace tensor_operation { ...@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i ...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i ...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_i
{ {
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmDefault>{}); device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances<GemmDefault>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instance.cpp
View file @ 363b6744
...@@ -8,7 +8,7 @@ namespace tensor_operation { ...@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_ ...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_ ...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_
{ {
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmKPadding>{}); device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances<GemmKPadding>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instance.cpp
View file @ 363b6744
...@@ -8,7 +8,7 @@ namespace tensor_operation { ...@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin ...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin ...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpaddin
{ {
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmMNKPadding>{}); device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances<GemmMNKPadding>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instance.cpp
View file @ 363b6744
...@@ -8,7 +8,7 @@ namespace tensor_operation { ...@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding ...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding ...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding
{ {
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_instances<GemmMNPadding>{}); device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_comp_instances<GemmMNPadding>{});
} }
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instance.cpp
View file @ 363b6744
...@@ -8,7 +8,7 @@ namespace tensor_operation { ...@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default ...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default ...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default
{ {
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances<Intrawave, device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances<Intrawave,
GemmDefault>{}); GemmDefault>{});
} }
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instance.cpp
View file @ 363b6744
...@@ -8,7 +8,7 @@ namespace tensor_operation { ...@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin ...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin ...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpaddin
{ {
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances<Intrawave, device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances<Intrawave,
GemmKPadding>{}); GemmKPadding>{});
} }
......
library/src/tensor_operation_instance/gpu/gemm_ab_scale/device_gemm_ab_scale_xdl_f8_f8_bf16/device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instance.cpp
View file @ 363b6744
...@@ -8,7 +8,7 @@ namespace tensor_operation { ...@@ -8,7 +8,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances( void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row, std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
Col, Col,
Tuple<>, Tuple<>,
...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd ...@@ -19,7 +19,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd
F32, F32,
Tuple<>, Tuple<>,
BF16, BF16,
128, 1,
128, 128,
128, 128,
PassThrough, PassThrough,
...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd ...@@ -28,7 +28,7 @@ void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadd
{ {
add_device_operation_instances( add_device_operation_instances(
instances, instances,
device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_instances<Intrawave, device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_1_128_128_mem_instances<Intrawave,
GemmMNKPadding>{}); GemmMNKPadding>{});
} }
......
profiler/src/CMakeLists.txt
View file @ 363b6744
# ckProfiler # ckProfiler
set(PROFILER_SOURCES set(PROFILER_SOURCES
profiler.cpp profiler.cpp
profile_gemm.cpp # profile_gemm.cpp
profile_reduce.cpp # profile_reduce.cpp
profile_groupnorm_bwd_data.cpp # profile_groupnorm_bwd_data.cpp
profile_groupnorm_fwd.cpp # profile_groupnorm_fwd.cpp
profile_layernorm_bwd_data.cpp # profile_layernorm_bwd_data.cpp
profile_layernorm_bwd_gamma_beta.cpp # profile_layernorm_bwd_gamma_beta.cpp
profile_groupnorm_bwd_gamma_beta.cpp # profile_groupnorm_bwd_gamma_beta.cpp
profile_layernorm_fwd.cpp # profile_layernorm_fwd.cpp
profile_max_pool2d_fwd.cpp # profile_max_pool2d_fwd.cpp
profile_pool3d_fwd.cpp # profile_pool3d_fwd.cpp
profile_avg_pool3d_bwd.cpp # profile_avg_pool3d_bwd.cpp
profile_max_pool3d_bwd.cpp # profile_max_pool3d_bwd.cpp
profile_avg_pool2d_bwd.cpp # profile_avg_pool2d_bwd.cpp
profile_max_pool2d_bwd.cpp # profile_max_pool2d_bwd.cpp
profile_softmax.cpp # profile_softmax.cpp
profile_batchnorm_fwd.cpp # profile_batchnorm_fwd.cpp
profile_batchnorm_bwd.cpp # profile_batchnorm_bwd.cpp
profile_batchnorm_infer.cpp # profile_batchnorm_infer.cpp
profile_conv_tensor_rearrange.cpp # profile_conv_tensor_rearrange.cpp
profile_transpose.cpp # profile_transpose.cpp
profile_permute_scale.cpp # profile_permute_scale.cpp
) )
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) # list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) # list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
endif() # endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
endif() # endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
endif() endif()
list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) # list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) # list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) # list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) # list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
endif() # endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
endif() # endif()
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
endif() # endif()
if(DL_KERNELS) # if(DL_KERNELS)
list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
endif() # endif()
set(PROFILER_EXECUTABLE ckProfiler) set(PROFILER_EXECUTABLE ckProfiler)
...@@ -93,89 +93,89 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) ...@@ -93,89 +93,89 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
target_compile_options(${PROFILER_EXECUTABLE} PRIVATE --offload-compress) target_compile_options(${PROFILER_EXECUTABLE} PRIVATE --offload-compress)
endif() endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
endif() # endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
endif() # endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
endif() endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
endif() # endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
endif() # endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif() # endif()
if(DL_KERNELS) # if(DL_KERNELS)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif() # endif()
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
profiler/src/profile_gemm_ab_scale.cpp
View file @ 363b6744
...@@ -32,8 +32,10 @@ enum struct GemmDataType ...@@ -32,8 +32,10 @@ enum struct GemmDataType
enum struct ScaleBlockTile enum struct ScaleBlockTile
{ {
Tile_128_128_128, // 0 Tile_128_128_128, // 0
Tile_1_128_128, // 1
}; };
#define OP_NAME "gemm_ab_scale" #define OP_NAME "gemm_ab_scale"
#define OP_DESC "GEMM_AB_Scale" #define OP_DESC "GEMM_AB_Scale"
...@@ -154,8 +156,25 @@ int profile_gemm_ab_scale(int argc, char* argv[]) ...@@ -154,8 +156,25 @@ int profile_gemm_ab_scale(int argc, char* argv[])
return pass ? 0 : 1; return pass ? 0 : 1;
}; };
// if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN &&
// scale_block_tile == ScaleBlockTile::Tile_128_128_128)
// {
// return profile(F8{},
// F32{},
// F8{},
// F32{},
// F8{},
// F32{},
// BF16{},
// ck::Number<128>{},
// ck::Number<128>{},
// ck::Number<128>{},
// Row{},
// Col{},
// Row{});
// }
if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN && if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN &&
scale_block_tile == ScaleBlockTile::Tile_128_128_128) scale_block_tile == ScaleBlockTile::Tile_1_128_128)
{ {
return profile(F8{}, return profile(F8{},
F32{}, F32{},
...@@ -164,7 +183,7 @@ int profile_gemm_ab_scale(int argc, char* argv[]) ...@@ -164,7 +183,7 @@ int profile_gemm_ab_scale(int argc, char* argv[])
F8{}, F8{},
F32{}, F32{},
BF16{}, BF16{},
ck::Number<128>{}, ck::Number<1>{},
ck::Number<128>{}, ck::Number<128>{},
ck::Number<128>{}, ck::Number<128>{},
Row{}, Row{},
......
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