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

CUTLASS 3.6.0 (#1850)

Browse Source 
* v3.6

* update changelog

* update readme

* fix typo

* fixing typos

* hopper gemm with weight prefetch

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Yujia Zhai
2024-10-09 15:33:27 -04:00
committed by GitHub
co-authored by yuzhai Haicheng Wu
parent 0837a2a00a
commit cc3c29a81a
354 changed files with 105937 additions and 8197 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cutlass/gemm/device/default_gemm_configuration.h
+52 -54
View File
@@ -764,6 +764,58 @@ struct DefaultGemmConfiguration<
////////////////////////////////////////////////////////////////////////////////
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;
};
////////////////////////////////////////////////////////////////////////////////
/// Base configuration for all {fe4m3, fe5m2} x {fe4m3, fe5m2} combinations on SM89
template <
typename ElementA,
@@ -793,60 +845,6 @@ 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<