ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

add int8 gemm multiply multiply a8w8 (#1591)

Browse Source 
* add int8 gemm multiply multiply a8w8

* uncomment

* clang-format-12

* Add example_gemm_multiply_multiply_xdl_int8

* Remove shell scripts

* update preprocess number for mi308; bring back printout in ckprofiler

* format

---------

Co-authored-by: chenjun <junchen2@amd.com>
Co-authored-by: Haocong WANG <haocwang@amd.com>
Co-authored-by: carlushuang <carlus.huang@amd.com>
This commit is contained in:
valarLip
2024-10-26 16:39:34 +08:00
committed by GitHub
parent eda5938386
commit 37f7afed1e
16 changed files with 794 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

105
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp
View File

@@ -96,6 +96,87 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_i
MultiplyMultiply>>>& instances);
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_INT8))
void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
I8,
I8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
I8,
I8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
I8,
I8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
I8,
I8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
I8,
I8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
Col,
Tuple<Row, Col>,
Row,
I8,
I8,
Tuple<F32, F32>,
BF16,
PassThrough,
PassThrough,
MultiplyMultiply>>>& instances);
#endif
template <typename ADataType,
typename BDataType,
typename CDataType,
@@ -155,6 +236,30 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
op_ptrs);
}
}
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_INT8))
if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}