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

v4.1 release update v2. (#2481)

Browse Source
This commit is contained in:
Junkai-Wu
2025-07-22 10:03:55 +08:00
committed by GitHub
parent 9baa06dd57
commit fd6cfe1ed0
179 changed files with 7878 additions and 1286 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/13_two_tensor_op_fusion/device/b2b_implicit_gemm_convolution.h
View File

@@ -177,7 +177,7 @@ public:
if(args.split_k_mode == SplitKMode::kParallel) {
// Split-K parallel: CTAs in k-dimension write the partial results in a temporary workspace.
// The user needs to call a reduction operator to optain the final output tensor
// The user needs to call a reduction operator to obtain the final output tensor
workspace_bytes =
sizeof(ElementAccumulator) *
size_t(cutlass::conv::implicit_gemm_tensor_c_size(kConvolutionalOperator, args.problem_size_0)) *

2
examples/13_two_tensor_op_fusion/fused_two_gemms_grouped_f16_sm80_rf.cu
View File

@@ -153,7 +153,7 @@ struct Options {
out << "13_fused_two_gemms_grouped_f16_sm80_rf\n\n"
<< " This example runs a grouped back-to-back GEMM kernel. A group of independent back-to-back GEMMs are\n"
<< " run in a single kernel. Each indivdual problem in the group is subject to the same constraints that non-grouped\n"
<< " run in a single kernel. Each individual problem in the group is subject to the same constraints that non-grouped\n"
<< " back-to-back GEMMs are subject to.s"
<< "Options:\n\n"
<< " --help If specified, displays this usage statement.\n\n"

6
examples/13_two_tensor_op_fusion/kernel/b2b_gemm.h
View File

@@ -248,7 +248,7 @@ struct B2bGemm {
typename Epilogue::OutputTileIterator::TensorRef* ref_C1;
typename Epilogue::OutputTileIterator::TensorRef* ref_D1;
// Epilogue params remain constant across all problmes in the group. Thus,
// Epilogue params remain constant across all problems in the group. Thus,
// the parameter here is not a pointer.
typename OutputOp0::Params epilogue0;
typename OutputOp1::Params epilogue1;
@@ -402,7 +402,7 @@ struct B2bGemm {
typename Epilogue::OutputTileIterator::TensorRef* ref_C1;
typename Epilogue::OutputTileIterator::TensorRef* ref_D1;
// Epilogue params remain constant across all problmes in the group. Thus,
// Epilogue params remain constant across all problems in the group. Thus,
// the parameter here is not a pointer.
typename OutputOp0::Params output_op_0;
typename OutputOp1::Params output_op_1;
@@ -434,7 +434,7 @@ struct B2bGemm {
// Only row-major outputs are currently supported, so no transpose is performed
}
/// Returns non-grouped paramaters to be used as input to the kernel-level
/// Returns non-grouped parameters to be used as input to the kernel-level
/// operator for the problem indicated by problem_visitor.
CUTLASS_HOST_DEVICE
Params to_single_params(const ProblemVisitor& problem_visitor) const {

2
examples/13_two_tensor_op_fusion/kernel/default_b2b_conv2d_fprop_sm80.h
View File

@@ -560,7 +560,7 @@ struct DefaultB2bConv2dFprop <
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines a kernel for Conv2dFprop specialization for Optimzed IteratorAlgorithm and
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm and
// multistage pipeline with interleaved layout.
template <
typename ElementA,

2
examples/13_two_tensor_op_fusion/kernel/default_b2b_conv2d_fprop_smem_accumulator_sm80.h
View File

@@ -606,7 +606,7 @@ struct DefaultB2bConv2dFprop <
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines a kernel for Conv2dFprop specialization for Optimzed IteratorAlgorithm and
/// Defines a kernel for Conv2dFprop specialization for Optimized IteratorAlgorithm and
// multistage pipeline with interleaved layout.
/// Accumulator will be staged in shared memory.
template <

2
examples/13_two_tensor_op_fusion/threadblock/b2b_mma_pipelined.h
View File

@@ -277,7 +277,7 @@ public:
IteratorAccumulatorScaleBias iterator_A1_scale, ///< iterator over A1 operand scale vectors in global memory
IteratorAccumulatorScaleBias iterator_A1_bias, ///< iterator over A1 operand bias vectors in global memory
IteratorB1 iterator_B1, ///< iterator over B1 operand in global memory
FragmentC0 const &src_accum, ///< source accumualtor tile
FragmentC0 const &src_accum, ///< source accumulator tile
OutputOp output_op_0, ///< epilogue operation after 1st Gemm
TransformA0 transform_A0 = TransformA0(), ///< transformation applied to A0 fragment
TransformB0 transform_B0 = TransformB0(), ///< transformation applied to B0 fragment

2
examples/13_two_tensor_op_fusion/threadblock/b2b_mma_pipelined_smem_accumulator.h
View File

@@ -298,7 +298,7 @@ public:
IteratorAccumulatorScaleBias iterator_accum0_scale, ///< iterator over D0 scale vector in global memory
IteratorAccumulatorScaleBias iterator_accum0_bias, ///< iterator over D0 bias vector in global memory
IteratorB1 iterator_B1, ///< iterator over B1 operand in global memory
FragmentC0 const &src_accum, ///< source accumualtor tile
FragmentC0 const &src_accum, ///< source accumulator tile
OutputOp output_op_0, ///< epilogue operation after 1st Gemm
TransformA0 transform_A0 = TransformA0(), ///< transformation applied to A0 fragment
TransformB0 transform_B0 = TransformB0(), ///< transformation applied to B0 fragment

2
examples/13_two_tensor_op_fusion/threadblock/default_b2b_mma.h
View File

@@ -93,7 +93,7 @@ template <
typename InstructionShape_,
/// Number of stages used in the pipelined mainloop
int Stages,
/// Operation perfomed by GEMM
/// Operation performed by GEMM
typename Operator,
/// Epilogue output operator
typename EpilogueOutputOp,