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

CUTLASS 3.5.0 (#1411)

Browse Source
This commit is contained in:
Vijay Thakkar
2024-03-19 17:51:04 -04:00
committed by GitHub
parent ffa34e7075
commit 629f4653c3
468 changed files with 48729 additions and 7252 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/13_two_tensor_op_fusion/kernel/b2b_gemm.h
+56 -59
View File
@@ -157,35 +157,34 @@ struct B2bGemm {
// Data members
//
GemmUniversalMode mode;
GemmCoord problem_size_0;
GemmCoord problem_size_1;
typename B2bMma::IteratorA0::TensorRef ref_A0;
typename B2bMma::IteratorB0::TensorRef ref_B0;
typename Epilogue::OutputTileIterator::TensorRef ref_C0;
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Scale0;
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Bias0;
typename B2bMma::IteratorB1::TensorRef ref_B1;
typename Epilogue::OutputTileIterator::TensorRef ref_C1;
typename Epilogue::OutputTileIterator::TensorRef ref_D1;
int64_t batch_stride_A0;
int64_t batch_stride_B0;
int64_t batch_stride_B1;
int64_t batch_stride_C1;
int64_t batch_stride_D1;
int64_t batch_stride_Bias0;
int64_t batch_stride_Scale0;
typename OutputOp0::Params epilogue0;
typename OutputOp1::Params epilogue1;
int batch_count;
GemmUniversalMode mode = cutlass::gemm::GemmUniversalMode::kGemm;
GemmCoord problem_size_0{0,0,0};
GemmCoord problem_size_1{0,0,0};
typename B2bMma::IteratorA0::TensorRef ref_A0{};
typename B2bMma::IteratorB0::TensorRef ref_B0{};
typename Epilogue::OutputTileIterator::TensorRef ref_C0{};
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Scale0{};
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Bias0{};
typename B2bMma::IteratorB1::TensorRef ref_B1{};
typename Epilogue::OutputTileIterator::TensorRef ref_C1{};
typename Epilogue::OutputTileIterator::TensorRef ref_D1{};
int64_t batch_stride_A0{0};
int64_t batch_stride_B0{0};
int64_t batch_stride_B1{0};
int64_t batch_stride_C1{0};
int64_t batch_stride_D1{0};
int64_t batch_stride_Bias0{0};
int64_t batch_stride_Scale0{0};
typename OutputOp0::Params epilogue0 {};
typename OutputOp1::Params epilogue1 {};
int batch_count{1};
//
// Methods
//
/// Default ctor
CUTLASS_HOST_DEVICE
Arguments() : mode(mode), problem_size_0(0, 0, 0), problem_size_1(0, 0, 0), batch_count(1) {}
Arguments() = default;
/// Constructs an Arguments structure
CUTLASS_HOST_DEVICE
@@ -285,47 +284,45 @@ struct B2bGemm {
/// Parameters structure
struct Params {
cutlass::gemm::GemmUniversalMode mode;
cutlass::gemm::GemmCoord problem_size_0;
cutlass::gemm::GemmCoord problem_size_1;
cutlass::gemm::GemmCoord grid_tiled_shape;
int swizzle_log_tile;
typename B2bMma::IteratorA0::Params params_A0;
typename B2bMma::IteratorA0::TensorRef ref_A0;
typename B2bMma::IteratorB0::Params params_B0;
typename B2bMma::IteratorB0::TensorRef ref_B0;
typename Epilogue::OutputTileIterator::Params params_C0;
typename Epilogue::OutputTileIterator::TensorRef ref_C0;
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Scale0;
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Bias0;
typename B2bMma::IteratorB1::Params params_B1;
typename B2bMma::IteratorB1::TensorRef ref_B1;
typename Epilogue::OutputTileIterator::Params params_C1;
typename Epilogue::OutputTileIterator::TensorRef ref_C1;
typename Epilogue::OutputTileIterator::Params params_D1;
typename Epilogue::OutputTileIterator::TensorRef ref_D1;
typename OutputOp0::Params output_op_0;
typename OutputOp1::Params output_op_1;
int64_t batch_stride_A0;
int64_t batch_stride_B0;
int64_t batch_stride_B1;
int64_t batch_stride_C1;
int64_t batch_stride_D1;
int64_t batch_stride_Bias0;
int64_t batch_stride_Scale0;
int *semaphore;
int gemm_k_iterations_0;
int gemm_k_size_0;
int gemm_k_iterations_1;
int gemm_k_size_1;
cutlass::gemm::GemmUniversalMode mode = cutlass::gemm::GemmUniversalMode::kGemm;
cutlass::gemm::GemmCoord problem_size_0{};
cutlass::gemm::GemmCoord problem_size_1{};
cutlass::gemm::GemmCoord grid_tiled_shape{};
int swizzle_log_tile{0};
typename B2bMma::IteratorA0::Params params_A0{};
typename B2bMma::IteratorA0::TensorRef ref_A0{};
typename B2bMma::IteratorB0::Params params_B0{};
typename B2bMma::IteratorB0::TensorRef ref_B0{};
typename Epilogue::OutputTileIterator::Params params_C0{};
typename Epilogue::OutputTileIterator::TensorRef ref_C0{};
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Scale0{};
typename B2bMma::IteratorAccumulatorScaleBias::TensorRef ref_Bias0{};
typename B2bMma::IteratorB1::Params params_B1{};
typename B2bMma::IteratorB1::TensorRef ref_B1{};
typename Epilogue::OutputTileIterator::Params params_C1{};
typename Epilogue::OutputTileIterator::TensorRef ref_C1{};
typename Epilogue::OutputTileIterator::Params params_D1{};
typename Epilogue::OutputTileIterator::TensorRef ref_D1{};
typename OutputOp0::Params output_op_0{};
typename OutputOp1::Params output_op_1{};
int64_t batch_stride_A0{0};
int64_t batch_stride_B0{0};
int64_t batch_stride_B1{0};
int64_t batch_stride_C1{0};
int64_t batch_stride_D1{0};
int64_t batch_stride_Bias0{0};
int64_t batch_stride_Scale0{0};
int *semaphore = nullptr;
int gemm_k_iterations_0{0};
int gemm_k_size_0{0};
int gemm_k_iterations_1{0};
int gemm_k_size_1{0};
//
// Methods
//
CUTLASS_HOST_DEVICE
Params(): mode(mode), swizzle_log_tile(0), semaphore(0), gemm_k_iterations_0(0), gemm_k_size_0(0),
gemm_k_iterations_1(0), gemm_k_size_1(0) { }
Params() = default;
CUTLASS_HOST_DEVICE
Params(