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

Support for Mixed Input TensorOp (#1084)

Browse Source 
* Passing warp-level mixed input F16*(S8/U8) tests

* passing device-level mixed input F16*(S8/U8) tests

* add to profiler - I8 (111 TFLOPs), U (123 TFLOPs)

* fast numeric conversions (I8 = 132 TFLOPs, U8 = 148 TFLOPs)

* Speedup reference compilation (REVERT THIS COMMIT)

* wider_add.u32_packed_sub.f16x2 (I8 = 132TFLOP/s, U8 = 170 TFLOP/s)

* Improve s8->f16 cvt and support bf16*u8 @158 TFLOPs

* BF16 * S8 (142 TFLOPs)

* Handle mixed-input upcast on OperandA (Support [S8|U8]*[F16|BF16]

* rename OpMultiplyAddMixedInput to OpMultiplyAddMixedInputUpcast

* Add device-level test and profiler support for upcast on operand A

* Move shfl before the cvt and reduce #shfls by 1/2

* fix smem_usage calculation for mixed_input types

* uncomment the stuff (getting ready for merge)

* profiler changes and mixed-input reference

* mixed input reference are in a new file

* use platform instead of std

* comments and typo only

* Use CreateGemmOperator and delete CreateMixedInputGemmOperator

* copyright for new files

* rebase follow-up
This commit is contained in:
Manish Gupta
2023-09-27 08:18:30 -07:00
committed by GitHub
parent 5cd735c48e
commit 7d8317a63e
26 changed files with 2064 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
include/cutlass/arch/mma.h
View File

@@ -68,14 +68,24 @@ struct OpMultiplyAddFastF16 {};
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Tag indicating the input data types are mixed and the narrower type is
/// upcasted to the wider type
struct OpMultiplyAddMixedInputUpcast {};
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Tag indicating the input is converted to 2 (big and small) TF32 components
// Perform 3xTF32 or 4xTF32 for every F32 output element
struct OpMultiplyAddFastF32 {};
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Tag indicating the input is converted to 2 (big and small) TF32 components
// Perform 3xTF32 or 4xTF32 for every complex<F32> output element
struct OpMultiplyAddComplexFastF32 {};
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Helper for determining whether staged accumulation should be used for a given operator
template <typename Operator>
struct UseStagedAccumulation {