* Allow per-column bias in EpilogueTensorBroadcast
EpilogueTensorBroadcast only supports per-row vector broadcast, because
the bias stride is hardcoded.
It can easily support both if the bias stride is made conditional, and
the original behavior is maintained by defaulting to per-row.
* Add unit test for EpilogueTensorBroadcast with per-col bias
---------
Co-authored-by: Ali Hassani <ahassanijr@gmail.com>
Co-authored-by: Ali Hassani <ali@hippoml.com>
* Fix inline ptx escaping for predicates.
Prevents `error: invalid % escape in inline assembly string` when compiling with clang.
* More double-quoting.
Currently, the default constructor of
`PredicatedTileAccessIteratorParams` will invoke undefined behavior in
its invocation of the `initialize` function. Specifically, it will
attempt to read from the uninitialized variables
`desc.element_size_bits` and `desc.advance_rank`. This commit changes
the default constructors of both `*Params` and `*Desc` to
zero-initialize all uninitialized members.
* Remove unused variables
* Qualify calls to make_fragment_? from templated base class.
Fixes clang build error.
* Add missing `#include <cstdio>`
* Various changes to fix clang compile errors.
* More changes to fix clang build.
Remaining issues:
- `params` initializer of `CollectiveEpilogue`.
- `ops` initializer of `Sm90VisitorImplBase`.
- `__usAtomicCAS` needs to be added to clang upstream.
* Fix remaining clang build issues.
* Qualify `cute::rank()` calls.
* Qualify some more calls that are otherwise ambiguous between `cute` and `std` namespace.
* Double-escape special registers in inline asm.
* small change
---------
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
Followup to #1224.
A change in the stream-k threadblock swizzle ctor since 3.3 breaks
single source GEMM with fused epilogue and stream-k. Multi-source was
already corrected.
Co-authored-by: Ali Hassani <ahassanijr@gmail.com>
* Release 3.3.0
Adds support for mixed precision GEMMs On Hopper and Ampere
Adds support for < 16B aligned GEMMs on Hopper
Enhancements to EVT
Enhancements to Python interface
Enhancements to Sub-byte type handling in CuTe
Several other bug-fixes and performance improvements.
* minor doc update
* set kIsHeavy member variables
* correct kIsHeavy value for Tanh
* set kIsHeavy=false for HardSwish
---------
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
* 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
when I use cutlass::epilogue:🧵:LinearCombinationSigmoid, I encounter the this error:
cutlass/include/cutlass/array.h(1549): error: no operator "-" matches these operands
Moving operator "-" from line 1549 to 1548 can solve this error
* [WIP] GEMM StreamK w/ Fused Epilogue
* Adds Gemm Streamk with Fused Epilogue kernel level struct.
* Mostly based on Gemm with Fused Epilogue,
* Requires a new epilogue
* Work in progress
* [WIP] StreamK support for GemmUniversalWithBroadcast
* Just based off of how StreamK is allowed in GemmUniversal
* Untested and a work in progress
* Minor fixes
* [WIP] It compiles!
It is almost certainly incorrect, but we're past getting the templates
to match, so checkpointing.
* Correction to reference kernel
* Fix typo
* Added MSE measurement
* Switch back to reference kernel + host for loop
Still WIP. Now we're getting even a larger MSE, but it's both on
basic Split-K and Stream-K.
* Fix typos
* Fix broadcast vector + requested changes
* Comment typo
* Small int option and more
* Fix incorrect condition on source needed
* Requested changes
* I think I got it?
* Bias vector should be stride 0
* Two source added!
* Typos
* Merge examples
* Bring back vector row offset
Just to ensure consistency with universal gemm with fused epilogue
* Base arguments and params structs for StreamK
* StreamK epilogue with broadcast now inherits the original
* undo params_streamk_base.h
---------
Co-authored-by: Ali Hassani <ahassanijr@gmail.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
* Remove references to device-only intrinsics when compiling for host.
Currently, we attempt to use the `__device__`-only functions
`__cvta_generic_to_shared` and `__nvvm_get_smem_pointer` when compiling
`cute::cast_smem_ptr_to_uint` for the host on Clang. This results in a
compilation error, as expected. This commit changes the definition of
the `*_ACTIVATED` macros so that they are only true when `__CUDA_ARCH__`
is defined; that is, when compiling for the device.
Additionally, the declaration of `__nvvm_get_smem_pointer`
is currently only visible during the device compilation pass when
compiling with NVCC; this commit makes the declaration visible during
host compilation with the `__device__` annotation.
* Annotate cute::cast_smem_ptr_to_uint as device-only.
The implementation of `cute::cast_smem_ptr_to_uint` is currently an
unchecked failure on host code, and the only host implementation I can
think of -- casting a probably-64-bit pointer to 32 bits somehow --
doesn't make sense to implement. This commit marks this function as
device-only so that it can't be accidentally used on host code.
* small change
---------
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
* Make operator() const-correct and add missing static functions.
Currently, `*Converter::operator()` requires a mutable object to invoke,
and there are missing `static result_type convert(source_type const &
source)` overloads for certain partial specializations of `*Converter`
objects. This commit makes `operator()` const-correct and adds missing
function overloads where appropriate.
* minor changes
* format
---------
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
- clang 1.14 complains about missing function from a host call:
cutlass/include/cute/arch/util.hpp:106:32: error: no matching function for call to '__cvta_generic_to_shared'
return static_cast<uint32_t>(__cvta_generic_to_shared(ptr));
- fixes this by defining CUTE_HOST_DEVICE for clang as well
Signed-off-by: Janusz Lisiecki <jlisiecki@nvidia.com>
Currently, the `LinearCombinationClamp` header file is not standalone,
and must have the definition of `cutlass::epilogue:🧵:ScaleType`
already available when it is `#include`d.
* Enable shared memory intrinsics and ldmatrix PTX on Clang.
This commit adds preprocessor checks to enable the shared memory
intrinsics `__cvta_generic_to_shared` and `__nvvm_get_smem_pointer`, as
well as the `ldmatrix` PTX instructions, on Clang. Preventing these
intrinsics from being used is a significant latency regression on Clang.
* refine the macro
---------
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
* Changes to iterators to support s8 gemm with f16 outputs
* should work
---------
Co-authored-by: Sujan Gonugondla <gsujan@amaon.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
Work around a likely GCC 8.x issue with fold expressions
and generic lambdas.
Only use the work-around when the host compiler is GCC 8.x.
This avoids any concerns about the work-around possibly
hindering inlining for a critical CuTe function (product).
Users can experiment with the work-around for other compilers
or compiler versions by defining the following macro.
CUTE_FOLD_GENERIC_LAMBDA_WORKAROUND
Fixes https://github.com/NVIDIA/cutlass/issues/788
Co-authored-by: Mark Hoemmen <mhoemmen@nvidia.com>
This commit changes the declarations of MMA operator class (SIMT, Tensor Core, WMMA Tensor Core) and operator type (multiply-add and so on) to definitions. This is done so that these tag structs are no longer incomplete types, which allows the `typeid` operator to be used on these tag structs. This is necessary for these tag structs to be used as type parameters in [GoogleTest typed tests](https://google.github.io/googletest/advanced.html#typed-tests).
This commit adds two `#include` directives so that the definitions of `cutlass::gemm::warp::WarpSize` from "cutlass/gemm/warp/mma.h" and `cutlass::arch::OpClassSimt` from "cutlass/arch/mma.h" are visible to "cutlass/epilogue/threadblock/default_epilogue_simt.h". Without them, there are compiler errors when building the header standalone:
```
In file included from cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.cu:1:
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:351:32: error: no member named 'warp' in namespace 'cutlass::gemm'; did you mean simply 'warp'?
static int const kWarpSize = cutlass::gemm::warp::WarpSize<arch::OpClassSimt>::value;
^
./cutlass/include/cutlass/epilogue/warp/tile_iterator_simt.h:49:11: note: 'warp' declared here
namespace warp {
^
In file included from cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.cu:1:
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:351:53: error: no member named 'WarpSize' in namespace 'cutlass::epilogue::warp'
static int const kWarpSize = cutlass::gemm::warp::WarpSize<arch::OpClassSimt>::value;
~~~~~~^
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:351:68: error: no member named 'OpClassSimt' in namespace 'cutlass::arch'
static int const kWarpSize = cutlass::gemm::warp::WarpSize<arch::OpClassSimt>::value;
~~~~~~^
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:351:82: error: no member named 'value' in the global namespace
static int const kWarpSize = cutlass::gemm::warp::WarpSize<arch::OpClassSimt>::value;
~~^
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:367:5: error: use of class template 'OutputTileThreadMap' requires template arguments
OutputTileThreadMap,
^
./cutlass/include/cutlass/epilogue/threadblock/output_tile_thread_map.h:134:8: note: template is declared here
struct OutputTileThreadMap : public OutputTileThreadMapHelpers<Iterations_, Delta_> {
^
In file included from cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.cu:1:
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:391:5: error: use of class template 'OutputTileThreadMap' requires template arguments
OutputTileThreadMap,
^
./cutlass/include/cutlass/epilogue/threadblock/output_tile_thread_map.h:134:8: note: template is declared here
struct OutputTileThreadMap : public OutputTileThreadMapHelpers<Iterations_, Delta_> {
^
In file included from cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.cu:1:
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:405:5: error: unknown type name 'OutputTileIterator'; did you mean 'WarpTileIterator'?
OutputTileIterator,
^
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:380:9: note: 'WarpTileIterator' declared here
using WarpTileIterator = cutlass::epilogue::warp::TileIteratorSimtDirect2dConv<
^
./cutlass/include/cutlass/epilogue/threadblock/default_epilogue_simt.h:408:5: error: use of class template 'SharedLoadIterator' requires template arguments
SharedLoadIterator,
^
./cutlass/include/cutlass/epilogue/threadblock/shared_load_iterator.h:67:7: note: template is declared here
class SharedLoadIterator {
^
```
* Relax stream K gemm alignment constraints
The current alignment requirements are too strict. Make them identical
to the checks for the regular universal gemm.
* Revert "Relax stream K gemm alignment constraints"
This reverts commit 31e80a250e2b0ac4bda2e4b437b39dc5bcd5e845.
* Relax stream K gemm alignment constraints
The current alignment requirements are too strict. Make them identical
to the checks for the regular universal gemm.
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
