MCPcopy Create free account

hub / github.com/tile-ai/tilelang / functions

Functions7,254 in github.com/tile-ai/tilelang

↓ 15 callersFunctionprim_expr_equal
Robust equality for PrimExpr shapes/extents. Tries structural_equal first, then falls back to expr_deep_equal. Python ints are converted
tilelang/utils/language.py:465
↓ 15 callersFunctionretrieve_ptr
Retrieve a pointer to the start of a (possibly sliced) buffer region. - Buffer -> base pointer - BufferRegion -> pointer with byte offse
tilelang/utils/language.py:317
↓ 15 callersFunctionto_tile_region
Encode a buffer-like object as a tl.region call for tile-op arguments.
tilelang/utils/language.py:237
↓ 14 callersFunction_check
(original, transformed)
testing/python/transform/test_tilelang_transform_inject_fence_proxy.py:12
↓ 14 callersMethod_safe_write_file
(path: str, mode: str, operation: Callable)
tilelang/cache/kernel_cache.py:449
↓ 14 callersFunctionassert_iter_map_simplify
(expect_dict, dom_map, predicate=True, check_level="surjective", simplify_trivial_iterators=True)
testing/python/arith/test_arith_iter_affine_map.py:85
↓ 14 callersFunctioncall_intrin
Build expression by calling an intrinsic function. Intrinsics can be overloaded with multiple data types via the intrinsic translation rule.
tilelang/language/tir/op.py:134
↓ 14 callersFunctiondetermine_fp8_type
Select the correct FP8 dtype string for the current platform. - CUDA defaults to FP8 E4M3FN / E5M2. - ROCm uses FNUZ except gfx950 (OCP),
tilelang/utils/target.py:152
↓ 14 callersMethodget_tile
(self, func)
tilelang/carver/roller/hint.py:97
↓ 14 callersFunctionkernel
( A, B, block_M=None, block_N=None, block_K=None, num_stages=None, thread_num=None
examples/analyze/example_gemm_analyze.py:12
↓ 14 callersMethodmake_mma_store_layout
Create a layout function for storing MMA results into a fragment buffer. This layout is used in conjunction with `inverse_mma_store_l
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:797
↓ 14 callersMethodrun
(self, code: str, timeout: float = 5.0)
tilelang/autodd.py:867
↓ 14 callersFunctionrun_gemm
( M, N, K, trans_A, trans_B, in_dtype, out_dtype, dtypeAccum, block_M,
testing/python/kernel/test_tilelang_kernel_gemm.py:52
↓ 14 callersFunctiontorch_assert_close
Custom function to assert that two tensors are "close enough," allowing a specified percentage of mismatched elements. Parameters: -
tilelang/utils/tensor.py:205
↓ 14 callersMethodvisit_stmt
(self, stmt: Stmt)
tilelang/analysis/ast_printer.py:36
↓ 13 callersMethod__init__
(self, args: ModelArgs)
examples/deepseek_v32/inference/model.py:542
↓ 13 callersFunction_check
(original, transformed)
testing/python/transform/test_tilelang_transform_loop_unswitching.py:8
↓ 13 callersFunction_make_ptx_mma
Factory to create a @dsl_user_op for ``mma.sync.aligned``. Args: ptx_shape: PTX shape, e.g. "m16n8k16" ptx_dtypes: PTX dtype suff
tilelang/contrib/cutedsl/ptx_mma.py:64
↓ 13 callersFunction_tl_math_op
(func, fastmath: bool, *args, **kwargs)
tilelang/contrib/cutedsl/math.py:56
↓ 13 callersMethodextract_thread_binding
is_m_first: True if the thread binding is in the form of (tx, warp_n, warp_m) which represents [warp_size, block_row_warps (split n),
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:240
↓ 13 callersMethodget_host_source
(self)
testing/python/autotune/test_tilelang_autotune_atomic_save.py:28
↓ 13 callersFunctionkernel
( A: T.Buffer(A_shape, in_dtype), B: T.Buffer(B_shape, storage_dtype), C: T.Buffer(C_s
examples/bitnet-1.58b/kernel_benchmark/tilelang_bitnet_158_int8xint2_decode.py:104
↓ 13 callersFunctionprepare_chunk_indices
( cu_seqlens: torch.LongTensor, chunk_size: int, )
examples/kda/FLA_KDA/fla_utils.py:107
↓ 12 callersFunction_apply_passes
Apply the LowerLDGSTG pass and related lowering passes.
testing/python/transform/test_tilelang_transform_lower_ldgstg.py:17
↓ 12 callersFunction_count_calls
(func: tvm.tirx.PrimFunc)
testing/python/transform/test_tilelang_transform_lower_ptx_async_copy.py:10
↓ 12 callersFunction_ensure_torch_dtypes
(*dtype_names)
maint/gemm/correctness_evaluation.py:449
↓ 12 callersMethod_get_cache_path
Gets the filesystem path for a cached kernel. Args: key (str): The hash key identifying the kernel. Returns:
tilelang/cache/kernel_cache.py:411
↓ 12 callersMethodcheck_continue_break
(self)
tilelang/language/eager/builder.py:282
↓ 12 callersFunctioncopy_impl
(A, B)
testing/python/language/test_tilelang_language_eager_jit.py:96
↓ 12 callersFunctionget_extent
Return the inferred extent (shape) of a buffer-like object. If `data` is a Var bound to a let value, the let value is resolved before inspection.
tilelang/language/utils.py:137
↓ 12 callersMethodget_thread_binding
(self)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:217
↓ 12 callersFunctionget_tvm_dtype
(value: AnyDType)
tilelang/language/dtypes.py:343
↓ 12 callersMethodldmatrix_b
(self, B_local_buf, B_shared_buf, ki, rk=0)
tilelang/rocm/intrinsics/wmma_macro_generator.py:229
↓ 12 callersMethodlower
( self, layout_map: dict, target: Target, thread_bounds: Range, thread
tilelang/cuda/op/gemm/gemm_mma.py:71
↓ 12 callersMethodswizzle_byte_size
(self)
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:71
↓ 11 callersMethod__init__
(self, config)
examples/bitnet-1.58b/modeling_bitnet.py:195
↓ 11 callersFunction_get_buffer_info
Extract buffer, shape, and dtype from BufferLikeType. Args: buffer_or_load_or_region: BufferLikeType Returns: tuple: (b
tilelang/layout/swizzle.py:12
↓ 11 callersMethodenable_cache
(self)
tilelang/env.py:358
↓ 11 callersMethodgenerate
( self, prompts: List[str], images: Optional[List[Image.Image]] = None, **kwar
examples/bitnet-1.58b/vllm_workspace/conftest.py:214
↓ 11 callersFunctionget_abs_err
(y, x)
examples/dsa_sparse_finetune/utils.py:4
↓ 11 callersFunctionget_err_ratio
(y, x)
examples/dsa_sparse_finetune/utils.py:10
↓ 11 callersMethodget_space_dim
(self)
tilelang/carver/roller/node.py:301
↓ 11 callersMethodldmatrix_a
(self, A_local_buf, A_shared_buf, ki, rk=0)
tilelang/rocm/intrinsics/wmma_macro_generator.py:175
↓ 11 callersFunctionprepare_token_indices
(cu_seqlens: torch.LongTensor)
examples/dsa_hisa/tilelang_utils.py:79
↓ 10 callersFunction__shfl_down_sync
Shuffle down within warp. Uses CuTeDSL's shuffle_sync_down with proper mask_and_clamp calculation. Matches CUDA: c = ((warpSize - width)
tilelang/contrib/cutedsl/warp.py:51
↓ 10 callersFunction__shfl_up_sync
Shuffle up within warp. Uses CuTeDSL's shuffle_sync_up with proper mask_and_clamp calculation. Matches CUDA: c = (warpSize - width) << 8
tilelang/contrib/cutedsl/warp.py:62
↓ 10 callersFunction_compile_tvm_ffi
(func, pass_configs, **kwargs)
testing/python/issue/test_tilelang_issue_tma_no_ws.py:11
↓ 10 callersMethod_load_cached_text_source
(self, source_attr: str, path_attr: str)
tilelang/jit/adapter/base.py:125
↓ 10 callersFunction_make_ptx_mma_sp
Factory to create a @dsl_user_op for ``mma.sp.sync.aligned``. Args: ptx_shape: PTX shape, e.g. "m16n8k32" ptx_dtypes: PTX dtype s
tilelang/contrib/cutedsl/ptx_mma.py:152
↓ 10 callersMethodapply
(self, target_labels: set[int])
tilelang/autodd.py:558
↓ 10 callersFunctionfn
()
testing/python/profiler/test_bench_cupti.py:12
↓ 10 callersFunctiongenerate_random_padding_mask
(max_seqlen, batch_size, device, mode="random")
examples/flash_attention/varlen_utils.py:7
↓ 10 callersFunctionget_e_factor
Return how many a_dtype elements are indexed by one meta_dtype element.
tilelang/utils/sparse.py:19
↓ 10 callersMethodget_input_buffers
(self)
tilelang/carver/roller/node.py:497
↓ 10 callersMethodget_input_shape
Get the input shape of the layout. Returns ------- List[int] The shape of the input layout.
tilelang/layout/layout.py:57
↓ 10 callersMethodis_equal
Check if the current fragment is equal to another fragment.
tilelang/layout/fragment.py:200
↓ 10 callersFunctionkernel
( block_M=None, block_N=None, block_K=None, num_stages=None, thread_nu
examples/gemm/example_gemm_autotune.py:117
↓ 10 callersFunctionparse_compute_version
Parse compute capability string to divide major and minor version Parameters ---------- compute_version : str compute capability
tilelang/contrib/nvcc.py:428
↓ 10 callersFunctionreduce
Perform a reduction operation on a buffer along a specified dimension. Args: buffer (tirx.Buffer): Input buffer to reduce out (ti
tilelang/language/reduce_op.py:24
↓ 10 callersFunctionretrieve_stride
Retrieve row-major strides for a buffer-like object based on its buffer.shape. For BufferRegion and BufferLoad, uses the underlying buffer's
tilelang/utils/language.py:267
↓ 10 callersFunctionrun_cumsum
(M, N, block_M, block_N, dim=0, reverse=False, dtype=T.float32, scope="smem")
testing/python/language/test_tilelang_language_scan.py:52
↓ 9 callersFunctionIsRDNA
()
examples/amd/example_amd_flash_attn_bwd.py:14
↓ 9 callersFunctionRange
Create a Range object. Parameters ---------- begin : PrimExpr The begin value of the range. end : Optional[PrimExpr]
tilelang/language/ast/ir.py:1709
↓ 9 callersFunction_as_uint32_mask
Normalize a warp lane mask to a uint32 TIR expression. Python literals (e.g. ``0xFFFFFFFF``) would otherwise be widened to int64 by TIR and p
tilelang/language/builtin.py:907
↓ 9 callersFunction_check
(original, transformed, target=auto_target)
testing/python/transform/test_tilelang_transform_pipeline_planning.py:13
↓ 9 callersFunction_collect_calls
(stmt, op_name: str)
testing/python/transform/test_tilelang_transform_fuse_mbarrier_arrive_expect_tx.py:19
↓ 9 callersMethod_cxx_cast
Render a C++ static_cast expression.
tilelang/jit/adapter/cutedsl/wrapper.py:833
↓ 9 callersFunction_get_buffer
Extract Buffer from Buffer, BufferLoad, or BufferRegion. Args: buffer_or_load_or_region: Can be Buffer, BufferLoad, or BufferRegion
tilelang/utils/language.py:16
↓ 9 callersFunction_get_let_stack
()
tilelang/language/frame.py:111
↓ 9 callersMethod_legalize_result_idx
(self, result_idx: list[int] | None)
tilelang/jit/adapter/base.py:33
↓ 9 callersFunction_lower_to_cuda_source
(func, target=SM80_TARGET)
testing/python/cuda/test_cuda_f32x2_intrinsics.py:95
↓ 9 callersFunction_tcgen05_ld_call
(tmem_ref, local_buf)
testing/python/transform/test_tilelang_transform_inject_tcgen05_fence.py:61
↓ 9 callersFunctionast_set_span
(ast: ast.AST, span: tuple[int, int, int, int])
tilelang/language/eager/ast.py:29
↓ 9 callersFunctionbuild_matmul_kernel
Compile and return a callable kernel that takes (A, B) and returns C.
maint/host_checks/common.py:27
↓ 9 callersMethodequivalent_function
Returns the function associated with this template. Returns: PrimFunc: The stored function.
tilelang/carver/template/base.py:108
↓ 9 callersMethodget_kernel_source
(self)
testing/python/cache/test_tilelang_kernel_cache_atomic_save.py:19
↓ 9 callersMethodget_tir
Retrieve a TIR (Tensor Intermediate Representation) PrimFunc from the stored callable or object.
tilelang/jit/__init__.py:303
↓ 9 callersFunctionkernel
( G: T.Tensor(G_shape, dtype=input_dtype), G_new: T.Tensor(G_shape, dtype=output_dtype), )
examples/gdn/example_cumsum.py:45
↓ 9 callersFunctionmaybe_contiguous
(x)
examples/attention_sink/example_mha_sink_bwd_bhsd.py:509
↓ 9 callersFunctionnum_splits_heuristic
Determines the optimal number of splits for maximizing GPU occupancy while balancing memory efficiency. Parameters: - total_mblocks (int
examples/blocksparse_attention/heuristic.py:4
↓ 9 callersFunctionpad_input
Arguments: hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask. indices: (total_nnz
examples/flash_attention/bert_padding.py:193
↓ 9 callersFunctionr
(op: type, i: int, m: OpMethod)
tilelang/language/parser/operation.py:113
↓ 9 callersMethodrecommend_hints
Provides a list of recommended hardware-aware configurations. Args: topk (int, optional): Number of top configurations t
tilelang/carver/template/base.py:153
↓ 9 callersFunctionregister_gemm_impl
Register a backend-specific GEMM implementation class.
tilelang/tileop/gemm/registry.py:23
↓ 9 callersFunctionrun_cummax
(M, N, block_M, block_N, dim=0, reverse=False, dtype=T.float32, scope="smem")
testing/python/language/test_tilelang_language_scan.py:366
↓ 9 callersFunctionset_autotune_inputs
Set input tensors for auto-tuning. This function creates a context manager for capturing input tensors during the auto-tuning process. It sup
tilelang/autotuner/capture.py:100
↓ 9 callersMethodsize
(self)
tilelang/carver/roller/bestfit.py:10
↓ 9 callersMethodwith_arch
Sets the architecture for this template and returns itself. Args: arch (TileDevice): The architecture to set. R
tilelang/carver/template/base.py:50
↓ 8 callersFunction_call_and_count_new_entries
Call fn and return how many *new* cache entries were added.
testing/python/autotune/test_tilelang_autotune_do_not_specialize.py:65
↓ 8 callersFunction_check
Apply DecoupleTypeCast pass and check IR matches expected output.
testing/python/transform/test_tilelang_transform_decouple_type_cast.py:8
↓ 8 callersFunction_collect_pipeline_loop_annotations
(func)
testing/python/transform/test_tilelang_transform_pipeline_planning.py:25
↓ 8 callersFunction_compile
(prim_func)
testing/python/language/test_tilelang_language_reduce_maxmin_nan.py:15
↓ 8 callersFunction_compile
(prim_func)
testing/python/language/test_tilelang_language_reduce.py:347
↓ 8 callersFunction_count_attrs
Count occurrences of a specific AttrStmt key in the function body.
testing/python/transform/test_tilelang_transform_lexical_alloc_scope.py:20
↓ 8 callersMethod_get_inputs
(self, with_output=False, dynamic_symbolic_constraints: dict[str, int] | None = None)
tilelang/profiler/__init__.py:62
↓ 8 callersFunction_legalize_dim
(buffer: tirx.Buffer, dim: int)
tilelang/language/reduce_op.py:12
↓ 8 callersFunction_make_reduce_kernel
(reduce_fn, length, dtype, *, nan_propagate)
testing/python/language/test_tilelang_language_reduce_maxmin_nan.py:19
↓ 8 callersFunction_run_pipeline_planning
(func, target=auto_target)
testing/python/transform/test_tilelang_transform_pipeline_planning.py:36
↓ 8 callersMethod_set_cached_text_source
(self, source_attr: str, path_attr: str, source: CachedTextSourceLike)
tilelang/jit/adapter/base.py:119
↓ 8 callersFunction_strip_block_reads_writes
Strip non-behavioral block metadata before structural comparison.
testing/python/transform/test_tilelang_transform_legalize_safe_memory_access.py:8
← previousnext →101–200 of 7,254, ranked by callers