When split-k is enabled, we should set alpha to 1 and beta to 0 for the
split-k gemm kernel.
The fix was from hwu36. I only did fixed some minor typos along with his
fix.
* Fix separate compilation `-dc`
- when cutlass is included in multiple compilation units
compiled with `-dc` OOB_NAN_F16x8 device constant is
instantiated multiple times causing
Multiple definition of '_ZN7cutlass4arch13OOB_NAN_F16x8E' error
This PR makes this variable a local constant as it is not
modified during runtime
Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
* Fix
Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
* Test GH
Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
* Revert test GH
Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
* Removed trivial copy constructors on parameter classes to enable device-side launch of CUTLASS kernels
* Added SFINAE to the `TensorRef(NonConstTensorRef const&)` constructor to avoid making it a copy-constructor for device code
* std => platform
* fix affine2
* really fix affine2
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
* Fix the build of cutlass/gemm/device/gemm_array.h and add a demo for GemmArray
* Add a reference to GemmArray to the docs
Co-authored-by: Ivan Komarov <dfyz@yandex-team.ru>
* Actually use float accumulation in gemm_f16t_f16t_f16t_wmma_tensor_op_f32_sm70.cu
As title
* Update gemm_f16t_f16t_f16t_wmma_tensor_op_f32_sm70.cu
change the missing one
Co-authored-by: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
* Support parallel split K mode for porfiling
Signed-off-by: Peter Han <fujun.han@iluvatar.ai>
* Parallel Split K support
1. find gemm kernel by preference key
2. switch m n for redution kernel
Signed-off-by: Peter Han <fujun.han@iluvatar.ai>
* parallel splitk for fp16 gemm
* add one missing file
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
* Add epilogue functor for residual block fusion
* Do not run split-k tests when ActivationOp is not Identity
* explain TestSplitK param
* return early
* Support half precision sigmoid activation
* introduce a vectorized variant using fast_tanh
* move the math to fast_math.h
* fixed compile
* .raw() -> .to_half()
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
* Support half precision sigmoid activation
* introduce a vectorized variant using fast_tanh
* refactored sigmoid using the new interface
* refactored gelu
* add silu activation
* add hardswish
* remove sigmoid for now
* add description to silu and hardswish, and other doc update
* Do not ignore Round
* use constant N
* Set isHeavy = true in sigmoid and silu epilogue
Updated GEMM performance plot with CUTLASS 2.8 compiled using CUDA 11.5 Toolkit.
GPUs under test:
NVIDIA A100
NVIDIA A2
NVIDIA TitanV
NVIDIA GeForce 2080 Ti
CUTLASS 2.7
Mainloop fusion for GEMM: summation over A or B
Strided DGRAD (optimized iterators)
Half-precision GELU_taylor activation functions
Use these when accumulation and epilogue compute types are all cutlass::half_t
Tuning and bug fixes to fused GEMM + GEMM example
Support for smaller than 128b aligned Convolutions: see examples
Caching of results to accelerate Convolution unit tests
Can be enabled or disabled by running cmake .. -DCUTLASS_TEST_ENABLE_CACHED_RESULTS=OFF
Corrections and bug fixes reported by the CUTLASS community
Thank you for filing these issues!
authored-by: Haicheng Wu haichengw@nvidia.com, Manish Gupta manigupta@nvidia.com, Dustyn Blasig dblasig@nvidia.com, Andrew Kerr akerr@nvidia.com
