leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 30 Packages Projects Releases Wiki Activity

Fix typos 2 (#842)

Browse Source 
Co-authored-by: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
This commit is contained in:
Alexander Pivovarov
2023-03-09 23:22:56 -05:00
committed by GitHub
co-authored by Haicheng Wu
parent c4f6b8c6bc
commit 7e370c9637
161 changed files with 310 additions and 309 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/library/scripts/generator.py
+2 -2
View File
@@ -89,7 +89,7 @@ def CreateGemmOperator(manifest, layouts, tile_descriptions, data_type, \
return operations
# Generates 3.0 API based GemmUniversal API kernels. Alignment constraits are folded in with layouts
# Generates 3.0 API based GemmUniversal API kernels. Alignment constraints are folded in with layouts
def CreateGemmUniversal3xOperator(
manifest, layouts, tile_descriptions, data_type,
complex_transforms=None,
@@ -4831,7 +4831,7 @@ if __name__ == "__main__":
parser.add_argument("--architectures", default='53;60;61;70;75;80', help="Target compute architectures")
parser.add_argument("--kernels", default='', help='Comma delimited list to filter kernels by name.')
parser.add_argument("--ignore-kernels", default='', help='Comma delimited list of kernels to exclude from build.')
parser.add_argument("--filter-by-cc", default='True', type=str, help='If enabled, kernels whose comupte capability range is not satisfied by the build target are excluded.')
parser.add_argument("--filter-by-cc", default='True', type=str, help='If enabled, kernels whose compute capability range is not satisfied by the build target are excluded.')
parser.add_argument("--cuda-version", default="11.0.0", help="Semantic version string of CUDA Toolkit")
parser.add_argument('--kernel-filter-file', type=str, default=None, required=False, help='Full path of filter file')
parser.add_argument('--selected-kernel-list', type=str, default=None, required=False,
tools/library/scripts/pycutlass/README.md
+2 -2
View File
@@ -83,7 +83,7 @@ The tiling size of above operations can also be customized.
### Using Docker
We recommend using one of our provided Docker images for using PyCUTLASS.
**To run CUTLASS 3 GEMM kernels targetting the NVIDIA Hopper architecture via PyCUTLASS,** you can use an included [Dockerfile](docker/Dockerfile-cuda12.0) based on the NGC CUDA 12.0 container:
**To run CUTLASS 3 GEMM kernels targeting the NVIDIA Hopper architecture via PyCUTLASS,** you can use an included [Dockerfile](docker/Dockerfile-cuda12.0) based on the NGC CUDA 12.0 container:
```shell
docker build -t pycutlass-cuda12.0:latest -f docker/Dockerfile-cuda12.0 .
docker run --gpus all -it --rm pycutlass-cuda12.0:latest
@@ -91,7 +91,7 @@ docker run --gpus all -it --rm pycutlass-cuda12.0:latest
Note that this Docker container does not include CuPy or PyTorch, and, thus, will not be able to run PyCUTLASS examples that
leverage these packages.
**To run CUTLASS 2.x kernels targetting pre-SM90 architectures via PyCUTLASS,** you can use an included [Dockerfile](docker/Dockerfile-cuda11.8-pytorch) based on an NGC PyTorch container:
**To run CUTLASS 2.x kernels targeting pre-SM90 architectures via PyCUTLASS,** you can use an included [Dockerfile](docker/Dockerfile-cuda11.8-pytorch) based on an NGC PyTorch container:
```shell
docker build -t pycutlass-cuda11.8-pytorch:latest -f docker/Dockerfile-cuda11.8-pytorch .
docker run --gpus all -it --rm pycutlass-cuda11.8-pytorch:latest
tools/library/scripts/pycutlass/docs/source/md/basic_idea.md
+4 -4
View File
@@ -10,7 +10,7 @@ PyCUTLASS handles the following things when launch the CUTLASS kernels
## Memory management
PyCUTLASS uses [RMM](https://github.com/rapidsai/rmm) to manage device memory. At the begining of the program, call
PyCUTLASS uses [RMM](https://github.com/rapidsai/rmm) to manage device memory. At the beginning of the program, call
```python
pycutlass.get_memory_pool({init_pool_size_in_bytes}, {max_pool_size_in_bytes})
```
@@ -41,7 +41,7 @@ The `{instruction_shape}` and `{opclass}` defines the instruction size and type.
|Opclass | element_a/element_b | element_acc | instruction_shape | math_operation |
| -- | -- | -- | -- | -- |
| cutlass.OpClass.TensorOp | cutlass.float64 | cutlass.float64 | [8, 8, 4] | MathOperation.multiply_add|
| | cutass.float32 cutlass.tfloat32, cutlass.float16 cutlass.bfloat16 | cutlass.float32 | [16, 8, 8] | MathOperation.multiply_add MathOperation.multiply_add_fast_f32 MathOperation.multiply_add_fast_f16 MathOperation.multiply_add_fast_bf16 |
| | cutlass.float32 cutlass.tfloat32, cutlass.float16 cutlass.bfloat16 | cutlass.float32 | [16, 8, 8] | MathOperation.multiply_add MathOperation.multiply_add_fast_f32 MathOperation.multiply_add_fast_f16 MathOperation.multiply_add_fast_bf16 |
| | cutlass.float16 | cutlass.float16/cutlass.float32|[16, 8, 16]| MathOperation.multiply_add |
| | cutlass.bfloat_16 | cutlass.float32 | [16, 8, 16]|MathOperation.multiply_add |
| | cutlass.int8 | cutlass.int32 | [16, 8, 32] | MathOperation.multiply_add_saturate|
@@ -178,7 +178,7 @@ import pycutlass
pycutlass.compiler.add_module([operation,])
```
Several operations can be compiled togather. The `nvcc` at `$CUDA_INSTALL_PATH/bin` is used by default as the compiler backend. But you can also switch to [CUDA Python](https://nvidia.github.io/cuda-python/overview.html)'s `nvrtc` with
Several operations can be compiled together. The `nvcc` at `$CUDA_INSTALL_PATH/bin` is used by default as the compiler backend. But you can also switch to [CUDA Python](https://nvidia.github.io/cuda-python/overview.html)'s `nvrtc` with
```python
pycutlass.compiler.nvrtc()
```
@@ -249,7 +249,7 @@ arguments = Conv2dArguments(
* `split_k_mode`: currently we support `cutlass.conv.SplitKMode.Serial` and `cutlass.conv.SplitKMode.Parallel`.
* `split_k_slice`: number of split-k slices
For ordianry conv2d, just use `cutlass.conv.SplitKMode.Serial` with `split_k_slice=1`.
For ordinary conv2d, just use `cutlass.conv.SplitKMode.Serial` with `split_k_slice=1`.
### Getting output_op
The way to create output_op is listed below
tools/library/scripts/pycutlass/src/cpp/include/arch.h
+3 -3
View File
@@ -51,9 +51,9 @@ void bind_opcode(py::module &m) {
.value("Simt", cutlass::OpcodeClass::kSimt,
R"pbdoc(Tag classifying math operators as thread-level operations)pbdoc")
.value("TensorOp", cutlass::OpcodeClass::kTensorOp,
R"pbdoc(Tag classifing operators as Tensor Core operations)pbdoc")
R"pbdoc(Tag classifying operators as Tensor Core operations)pbdoc")
.value("WmmaTensorOp", cutlass::OpcodeClass::kWmmaTensorOp,
R"pbdoc(Tag classifing operators as WMMA Tensor Core operations)pbdoc")
R"pbdoc(Tag classifying operators as WMMA Tensor Core operations)pbdoc")
.value("SparseTensorOp", cutlass::OpcodeClass::kSparseTensorOp,
R"pbdoc(Tag classifing operators as sparseTensor Core operations)pbdoc");
R"pbdoc(Tag classifying operators as sparseTensor Core operations)pbdoc");
}
tools/library/scripts/pycutlass/src/cpp/include/conv/conv_problem_size.h
+1 -1
View File
@@ -42,7 +42,7 @@ namespace py = pybind11;
void bind_conv_problem_size(py::module &m) {
//
// Conv2d Problem Size:
// include/cutlass/conv/conv2d_problem_sizd.h
// include/cutlass/conv/conv2d_problem_size.h
//
py::class_<cutlass::conv::Conv2dProblemSize>(m, "Conv2dProblemSize")
// constructors
tools/library/scripts/pycutlass/src/cpp/include/epilogue/epilogue_visitor_op/visitor_op_binary.h
+1 -1
View File
@@ -87,7 +87,7 @@ public:
using BinaryOp = BinaryOp_<ElementCompute, kElementsPerAccess>;
static_assert(kElementsPerAccess==VisitAccessTypeA::kElements, "kElementsPerAccess mismatches with Visitor A");
static_assert(kElementsPerAccess==VisitAccessTypeB::kElements, "kElementsPerAccess misnatches with Visitor B");
static_assert(kElementsPerAccess==VisitAccessTypeB::kElements, "kElementsPerAccess mismatches with Visitor B");
/// SMEM buffer class required in the epilogue visitor
struct SharedStorage {
tools/library/scripts/pycutlass/src/cpp/include/epilogue/epilogue_visitor_op/visitor_op_column_reduction.h
+2 -2
View File
@@ -55,7 +55,7 @@ template <
typename ElementReduction_, ///< Data type of the output reduction in device memory
typename ElementReductionAccumulator_ , ///< Data type to accumulate reduction in smem and register
typename OutputTileIterator_, ///< Tile Iterator type
typename Visitor_ ///< preceeding visitor op
typename Visitor_ ///< preceding visitor op
>
class VisitorOpColumnReduction {
public:
@@ -83,7 +83,7 @@ public:
/// Fragment type of accumulator
using AccumulatorAccessType = Array<ElementAccumulator, kElementsPerAccess>;
/// Fragment type of redcution
/// Fragment type of reduction
using ReductionAccumulatorAccessType = Array<ElementReductionAccumulator, kElementsPerAccess>;
/// Thread map used by output tile iterators
tools/library/scripts/pycutlass/src/cpp/include/epilogue/epilogue_visitor_op/visitor_op_linear_combination.h
+1 -1
View File
@@ -86,7 +86,7 @@ public:
using CombinationOp = cutlass::plus<VisitAccessType>;
static_assert(kElementsPerAccess==VisitAccessTypeA::kElements, "kElementsPerAccess mismatches with Visitor A");
static_assert(kElementsPerAccess==VisitAccessTypeB::kElements, "kElementsPerAccess misnatches with Visitor B");
static_assert(kElementsPerAccess==VisitAccessTypeB::kElements, "kElementsPerAccess mismatches with Visitor B");
/// SMEM buffer class required in the epilogue visitor
struct SharedStorage {
tools/library/scripts/pycutlass/src/cpp/include/epilogue/epilogue_visitor_op/visitor_op_row_reduction.h
+2 -2
View File
@@ -56,7 +56,7 @@ template <
typename ElementReduction_, ///< Data type of the output reduction in device memory
typename ElementReductionAccumulator_ , ///< Data type to accumulate reduction in smem and register
typename OutputTileIterator_, ///< Tile Iterator type
typename Visitor_ ///< preceeding visitor op
typename Visitor_ ///< preceding visitor op
>
class VisitorOpRowReduction {
public:
@@ -82,7 +82,7 @@ public:
/// Fragment type of accumulator
using AccumulatorAccessType = Array<ElementAccumulator, kElementsPerAccess>;
/// Fragment type of redcution
/// Fragment type of reduction
using ReductionAccumulatorAccessType = Array<ElementReductionAccumulator, kElementsPerAccess>;
/// Thread map used by output tile iterators
tools/library/scripts/pycutlass/src/cpp/include/gemm/gemm.h
+1 -1
View File
@@ -52,7 +52,7 @@ void bind_gemm(py::module &m) {
.value("Array", cutlass::gemm::GemmUniversalMode::kArray)
.value("Invalid", cutlass::gemm::GemmUniversalMode::kInvalid);
/// GemmCoord is a structure that specifies a location within the coordiate space of a GEMM problem
/// GemmCoord is a structure that specifies a location within the coordinate space of a GEMM problem
py::class_<cutlass::gemm::GemmCoord>(m, "GemmCoord")
.def(py::init<int, int, int>())
.def("m", py::overload_cast<>(&cutlass::gemm::GemmCoord::m))
tools/library/scripts/pycutlass/src/cpp/include/swizzling.h
+3 -3
View File
@@ -88,7 +88,7 @@ void bind_identity_swizzle(py::module & m, std::string name) {
R"pbdoc(Computes CUDA grid dimensions given a size in units of logical tiles)pbdoc")
.def("tag", [](const T & swizzle){
return demangle(typeid(T).name());
}, R"pbdoc(Returns the c++ name of the swizzling for code emittion)pbdoc");
}, R"pbdoc(Returns the c++ name of the swizzling for code emission)pbdoc");
}
template<typename T>
@@ -109,7 +109,7 @@ void bind_swizzle(py::module & m, std::string name, std::string doc) {
R"pbdoc(Computes CUDA grid dimensions given a size in units of logical tiles)pbdoc")
.def("tag", [](const T & swizzle){
return demangle(typeid(T).name());
}, R"pbdoc(Returns the c++ name of the swizzling for code emittion)pbdoc");
}, R"pbdoc(Returns the c++ name of the swizzling for code emission)pbdoc");
}
template<typename T>
@@ -132,7 +132,7 @@ void bind_dgrad_swizzle(py::module & m, std::string name) {
R"pbdoc(Computes CUDA grid dimensions given a size in units of logical tiles)pbdoc")
.def("tag", [](const T & swizzle){
return demangle(typeid(T).name());
}, R"pbdoc(Returns the c++ name of the swizzling for code emittion)pbdoc");
}, R"pbdoc(Returns the c++ name of the swizzling for code emission)pbdoc");
}
void bind_threadblock_swizzle(py::module &m) {
tools/library/scripts/pycutlass/src/cpp/include/tensor_ref_view.h
+1 -1
View File
@@ -20,7 +20,7 @@
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSE<cutlass::TensorRef<QUENTIAL
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
tools/library/scripts/pycutlass/src/pycutlass/builder/collective_op_builder.py
+4 -4
View File
@@ -97,7 +97,7 @@ def cluster_shape_to_tma(dim):
def make_cpasync_gmem_tiled_copy(thread_count, element, alignment, gmma_layout, dim_mn, dim_k):
"""
Returns a `make_tiled_copy` call for a given configuraiton
Returns a `make_tiled_copy` call for a given configuration
:param thread_count: number of threads in the threadblock
:type thread_count: int
@@ -296,7 +296,7 @@ def common_values(op, stage_count, transform_A, transform_B):
def build_gmma_tma(op):
"""
Builds a collective operation declaration targetting TMA GMMA kernels
Builds a collective operation declaration targeting TMA GMMA kernels
:param op: GEMM operation for which to build a collective operation
:type op: pycutlass.GemmOperation
@@ -335,7 +335,7 @@ def build_gmma_tma(op):
def build_gmma_cpasync(op):
"""
Builds a collective operation declaration targetting cp.async GMMA kernels
Builds a collective operation declaration targeting cp.async GMMA kernels
:param op: GEMM operation for which to build a collective operation
:type op: pycutlass.GemmOperation
@@ -377,7 +377,7 @@ def build_gmma_cpasync(op):
def build(operation):
"""
Builds a collective operation declaration targetting cp.async or TMA for GMMA kernels
Builds a collective operation declaration targeting cp.async or TMA for GMMA kernels
:param operation: GEMM operation for which to build a collective operation
:type operation: pycutlass.GemmOperation
tools/library/scripts/pycutlass/src/pycutlass/c_types.py
+1 -1
View File
@@ -150,7 +150,7 @@ def get_gemm_arguments(epilogue_functor):
("ldc", ctypes.c_longlong),
("ldd", ctypes.c_longlong),
("ptr_gather_A_indices", ctypes.c_void_p),
("ptr_gether_B_indices", ctypes.c_void_p),
("ptr_gather_B_indices", ctypes.c_void_p),
("ptr_scatter_D_indices", ctypes.c_void_p)
]
tools/library/scripts/pycutlass/src/pycutlass/gemm_operation.py
+8 -8
View File
@@ -278,7 +278,7 @@ class GemmArguments2x(ArgumentBase):
# get launch configuration
launch_config = self.operation.rt_module.plan(self)
# get the host and evice workspace
# get the host and device workspace
device_workspace_size = \
self.operation.rt_module.get_device_workspace_size(self)
@@ -293,7 +293,7 @@ class GemmArguments2x(ArgumentBase):
device_workspace = 0
if (workspace_ptr is not None and
self.gemm_mode == cutlass.gemm.Mode.GemmSplitKParallel):
# in GEMM splik-K parallel, the D pointer is redirected
# in GEMM split-K parallel, the D pointer is redirected
# to the workspace
self.ptr_D = cuda.CUdeviceptr(workspace_ptr)
elif (workspace_ptr is not None and
@@ -350,7 +350,7 @@ class GemmArguments3x(GemmArguments2x):
A: 'Tensor', B: 'Tensor', C: 'Tensor', D: 'Tensor',
gemm_mode: 'cutlass.gemm.Mode'=cutlass.gemm.Mode.Gemm, **kwargs):
if gemm_mode not in [cutlass.gemm.Mode.Gemm, cutlass.gemm.Mode.Batched]:
raise Exception("Unsupporged GEMM mode {}.".format(gemm_mode))
raise Exception("Unsupported GEMM mode {}.".format(gemm_mode))
super().__init__(operation, problem_size, A, B, C, D, gemm_mode, **kwargs)
@@ -387,7 +387,7 @@ class GemmArguments3x(GemmArguments2x):
)
def initialize(self):
# get the host and evice workspace
# get the host and device workspace
device_workspace_size = \
self.operation.rt_module.get_device_workspace_size(self)
@@ -402,7 +402,7 @@ class GemmArguments3x(GemmArguments2x):
device_workspace = 0
if (workspace_ptr is not None and
self.gemm_mode == cutlass.gemm.Mode.GemmSplitKParallel):
# in GEMM splik-K parallel, the D pointer is redirected
# in GEMM split-K parallel, the D pointer is redirected
# to the workspace
self.ptr_D = cuda.CUdeviceptr(workspace_ptr)
elif (workspace_ptr is not None and
@@ -610,7 +610,7 @@ class GemmGroupedArguments:
# get launch configuration
launch_config = self.operation.rt_module.plan(self)
# get the host and evice workspace
# get the host and device workspace
device_workspace_size = \
self.operation.rt_module.get_device_workspace_size(self)
@@ -1072,7 +1072,7 @@ class GemmOperationBase:
self.api = api
self.prefix = "3x" if self.api == ApiVersion.v3x else ""
# use deep copy to avoid overwritting the original TensorDescription
# use deep copy to avoid overwriting the original TensorDescription
if self.api != ApiVersion.v3x and C.layout == cutlass.ColumnMajor:
#: Operand A
self.A: TensorDescription = copy.deepcopy(B)
@@ -1300,7 +1300,7 @@ class GemmOperationGrouped(GemmOperationBase):
super(GemmOperationGrouped, self).__init__(GemmKind.Grouped, arch, tile_description,
A, B, C, epilogue_functor, swizzling_functor, **kwargs)
assert "precompute_mode" in kwargs.keys(
), "missing keyword arguement 'precompute_mode'."
), "missing keyword argument 'precompute_mode'."
self.precompute_mode = kwargs["precompute_mode"]
self.rt_module = GemmRTGrouped(self)
self.argument_type = self.rt_module.argument_type
tools/library/scripts/pycutlass/src/pycutlass/library.py
+1 -1
View File
@@ -713,7 +713,7 @@ class TileDescription:
"""
:param threadblock_shape: shape of a threadblock tyle
:type threadblock_shape: list or tuple
:param stages: number of pipline stages in the operation. For SM90 kernels, this can be set to `None` and the maximum
:param stages: number of pipeline stages in the operation. For SM90 kernels, this can be set to `None` and the maximum
number of stages that can be supported for an operation on a given architecture will be computed at a later time
:type stages: int or None
:param warp_count: number of warps in each [M, N, K] dimension of a threadblock tile
tools/library/scripts/pycutlass/src/pycutlass/reduction_operation.py
+2 -2
View File
@@ -108,7 +108,7 @@ class ReductionArguments:
if layout == cutlass.RowMajor:
return TensorRef2D_(int(device_ptr), extent[1])
else:
raise ValueError("unknonwn layout type")
raise ValueError("unknown layout type")
def get_arguments(self):
ref_workspace = ReductionArguments.get_tensor_ref(
@@ -320,7 +320,7 @@ class ReductionOperation:
#
def procedural_name(self):
''' The full procedural name indicates architeture, extended name, tile size'''
''' The full procedural name indicates architecture, extended name, tile size'''
return self.configuration_name()
def run(self, arguments: ReductionArguments) -> cuda.CUresult:
tools/library/scripts/pycutlass/src/pycutlass/test/conv2d_testbed.py
+1 -1
View File
@@ -498,7 +498,7 @@ class Conv2dLauncher:
########################################################################################################
# TestAllConv: Runs cutlass::conv::device::ImplicitGemmConvolution operator and compares it with reference
# TestAllConv runs conv operator on default conv problem sizes from test::conv::device::TestbedConv2dProblemSizes
# Additionaly, each conv2d test can provide conv problem sizes (conv_test_sizes) and blacklist of sizes
# Additionally, each conv2d test can provide conv problem sizes (conv_test_sizes) and blacklist of sizes
# (conv_blacklist_sizes)
############################################################################################################
tools/library/scripts/pycutlass/src/pycutlass/test/gemm_testbed.py
+1 -1
View File
@@ -489,7 +489,7 @@ def test_all_gemm(operation: 'GemmOperationUniversal', testcase="universal"):
else:
alignment = 128 // minimum_operand_element_size
# int8_t gemm alignment constrainst
# int8_t gemm alignment constraints
if opcode_class == cutlass.OpClass.Simt and operation.A.element == cutlass.int8 and operation.A.layout == cutlass.ColumnMajor:
alignment_m = 4
else:
tools/library/scripts/pycutlass/src/pycutlass/test/utils.py
+1 -1
View File
@@ -63,7 +63,7 @@ def get_name(layouts, alignments, element_output,
Generates a procedural name for a test case.
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alingments of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
tools/library/scripts/pycutlass/test/gemm/gemm_bf16_sm90.py
+1 -1
View File
@@ -52,7 +52,7 @@ def add_test(cls, layouts, alignments, element_output, element_accumulator, elem
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alingments of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
tools/library/scripts/pycutlass/test/gemm/gemm_f16_sm90.py
+2 -2
View File
@@ -42,7 +42,7 @@ from pycutlass.test.gemm_testbed import test_all_gemm
from pycutlass.utils.device import device_cc
# Partial specialziation for naming tests
# Partial specialization for naming tests
name_fn = partial(get_name, element_a=cutlass.float16, element_b=cutlass.float16, arch=90)
@@ -54,7 +54,7 @@ def add_test(cls, layouts, alignments, element_output, element_accumulator, elem
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alingments of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
tools/library/scripts/pycutlass/test/gemm/gemm_f64_sm90.py
+1 -1
View File
@@ -52,7 +52,7 @@ def add_test(cls, layouts, alignments, element_output, element_accumulator, elem
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alingments of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue
tools/library/scripts/pycutlass/test/gemm/gemm_s8_sm90.py
+1 -1
View File
@@ -52,7 +52,7 @@ def add_test(cls, layouts, alignments, element_output, element_accumulator, elem
:param cls: class to which the generated method will be added
:type cls: type
:param layouts: indexable container of layouts of A, B, and C operands
:param alignments: indexable container of alingments of A, B, and C operands
:param alignments: indexable container of alignments of A, B, and C operands
:param element_output: data type of the output element
:param element_accumulator: data type used in accumulation
:param element_epilogue: data type used in computing the epilogue