leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity

Add support for mixed 4-bit/8-bit data types GEMM (#1413)

Browse Source 
* Add support for mixed 4-bit/8-bit data types GEMM

* fix ( and )

---------

Co-authored-by: Aleksandar Samardžić <asamardzic@matf.bg.ac.rs>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Aleksandar Samardžić
2024-08-29 23:11:06 -04:00
committed by GitHub
co-authored by Aleksandar Samardžić Haicheng Wu
parent f7b19de32c
commit e1976daacc
15 changed files with 960 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cutlass/gemm/device/default_gemm_configuration.h
+54
View File
@@ -793,6 +793,60 @@ struct DefaultGemmConfigurationSm89F8 {
using Operator = arch::OpMultiplyAdd;
};
////////////////////////////////////////////////////////////////////////////////
template <
typename ElementC>
struct DefaultGemmConfiguration<
arch::OpClassTensorOp,
arch::Sm80,
int4b_t,
int8_t,
ElementC,
int32_t> {
static int const kAlignmentA = 128 / sizeof_bits<int4b_t>::value;
static int const kAlignmentB = 128 / sizeof_bits<int8_t>::value;
using ThreadblockShape = GemmShape<128, 256, 64>;
using WarpShape = GemmShape<64, 64, 64>;
using InstructionShape = GemmShape<16, 8, 32>;
static int const kStages = 3;
using EpilogueOutputOp = epilogue::thread::LinearCombinationClamp<
ElementC, 128 / sizeof_bits<ElementC>::value, int32_t, float>;
using Operator = arch::OpMultiplyAddSaturate;
};
////////////////////////////////////////////////////////////////////////////////
template <
typename ElementC>
struct DefaultGemmConfiguration<
arch::OpClassTensorOp,
arch::Sm80,
int8_t,
int4b_t,
ElementC,
int32_t> {
static int const kAlignmentA = 128 / sizeof_bits<int8_t>::value;
static int const kAlignmentB = 128 / sizeof_bits<int4b_t>::value;
using ThreadblockShape = GemmShape<128, 256, 64>;
using WarpShape = GemmShape<64, 64, 64>;
using InstructionShape = GemmShape<16, 8, 32>;
static int const kStages = 3;
using EpilogueOutputOp = epilogue::thread::LinearCombinationClamp<
ElementC, 128 / sizeof_bits<ElementC>::value, int32_t, float>;
using Operator = arch::OpMultiplyAddSaturate;
};
////////////////////////////////////////////////////////////////////////////////
/// Partial specialization for SM89 fe4m3 x fe4m3
template <typename ElementC, typename ElementAccumulator>
struct DefaultGemmConfiguration<