MCPcopy Create free account

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

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

↓ 8 callersFunction_to_ptr
(value, access_type: str)
tilelang/language/cluster.py:21
↓ 8 callersFunctionauto_infer_current_arch
()
tilelang/carver/arch/__init__.py:34
↓ 8 callersFunctioncast
Cast ``value`` to ``dtype`` with optional PTX-style rounding hints. Parameters ---------- value : object The source operand.
tilelang/language/tir/ir.py:331
↓ 8 callersFunctionchunk_local_cumsum
( g: torch.Tensor, chunk_size: int, reverse: bool = False, scale: float = None, cu_seqlens
examples/kda/FLA_KDA/cumsum.py:431
↓ 8 callersFunctioncompress
( A: torch.Tensor, meta_dtype: torch.dtype | None = None, # noqa: FA100 block_m: int | None = Non
tilelang/utils/sparse.py:251
↓ 8 callersMethodconstexpr
(self, name: str, dtype: str = "int32")
tilelang/language/eager/builder.py:754
↓ 8 callersFunctionevaluate
Evaluate the input expression. Parameters ---------- value: PrimExpr The input expression to evaluate.
tilelang/language/ast/ir.py:1316
↓ 8 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/rocm/intrinsics/mfma_macro_generator.py:345
↓ 8 callersMethodfind_frame_idx
(self, frame: type | tuple[type, ...], start=0)
tilelang/language/eager/builder.py:272
↓ 8 callersMethodfrom_code
(cls, kind: ASTPatKind, code: str, placeholders: set[str])
tilelang/autodd.py:294
↓ 8 callersFunctionget_device_attribute
(attr: int, device_id: int = 0)
tilelang/carver/arch/driver/cuda_driver.py:67
↓ 8 callersFunctionget_target_compute_version
Utility function to get compute capability of compilation target. Looks for the target arch in three different places, first in the target input,
tilelang/contrib/nvcc.py:387
↓ 8 callersFunctionis_cuda_target
(target: Target)
tilelang/jit/adapter/utils.py:97
↓ 8 callersMethodis_float4
(self)
tilelang/language/dtypes.py:22
↓ 8 callersMethodis_none
(self)
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:59
↓ 8 callersFunctionprepare_input
( B, S, H, DK, DV, chunk_size, input_dtype, output_dtype, accum_dtype,
examples/gdn/example_chunk_delta_bwd.py:32
↓ 8 callersFunctionquote1
(expr: str, *, passes: list[Any] | None = None, span=None, **kws)
tilelang/language/eager/ast.py:66
↓ 8 callersFunctionquote_expr
(expr: str, **kws)
tilelang/language/eager/ast.py:72
↓ 8 callersMethodrun_once
(self, func: Callable | None = None)
tilelang/jit/kernel.py:469
↓ 8 callersMethodswizzle_byte_size
(self)
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:68
↓ 8 callersFunctiontarget_get_mcpu
(target: str | Target | None)
tilelang/utils/target.py:40
↓ 8 callersFunctiontransform_func
(*args)
tilelang/cuda/intrinsics/layout/mma_layout.py:249
↓ 8 callersFunctionunpad_input
Arguments: hidden_states: (batch, seqlen, ...) attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
examples/flash_attention/bert_padding.py:94
↓ 7 callersFunctionCUDAPassPipelineBodyPrologue
(mod: IRModule, target: Target)
tilelang/cuda/pipeline.py:68
↓ 7 callersMethod__enter__
(self)
tilelang/language/eager/builder.py:114
↓ 7 callersMethod__new__
Implements singleton pattern for KernelCache class. Returns: KernelCache: The singleton instance of KernelCache.
tilelang/cache/kernel_cache.py:163
↓ 7 callersFunction_assert_legalize_matches_expected
(before, expected, strip_annotations: bool = False)
testing/python/transform/test_tilelang_transform_legalize_safe_memory_access.py:69
↓ 7 callersMethod_collect_host_kernel_call_sites
Collect CuTeDSL kernel calls from the lowered host entry in order.
tilelang/jit/adapter/cutedsl/wrapper.py:850
↓ 7 callersFunction_count_calls
(stmt, op_name: str)
testing/python/transform/test_tilelang_transform_inject_tcgen05_fence.py:29
↓ 7 callersFunction_emit_tmem_ld
Recursively split x-count into power-of-2 segments and emit TMEM loads. Called during @cute.jit compilation. n_x: remaining x-elem
tilelang/contrib/cutedsl/gemm_tcgen05.py:473
↓ 7 callersMethod_get_staging_root
()
tilelang/cache/kernel_cache.py:209
↓ 7 callersMethod_load_binary
Load binary content from a cache artifact.
tilelang/autotuner/param.py:164
↓ 7 callersFunction_mbar_to_buffer_load
Convert a memory barrier to a buffer load. Args: mbar: BarrierType The memory barrier to convert Returns: tirx.B
tilelang/language/builtin.py:34
↓ 7 callersMethod_pythonic_expr
(self, expr: tvm.tirx.PrimExpr)
tilelang/jit/adapter/wrapper.py:263
↓ 7 callersMethod_pythonic_expr
Convert TVM expression to Python string, ignoring casts. Casts are noise in generated Python code - Python is dynamically typed.
tilelang/jit/adapter/nvrtc/wrapper.py:284
↓ 7 callersFunction_read_cmake_bool
(i: str | None, default=False)
version_provider.py:21
↓ 7 callersMethod_save_kernel_to_disk
Persists a compiled kernel to disk cache using atomic directory rename. All files are first written into a temporary staging directo
tilelang/cache/kernel_cache.py:464
↓ 7 callersFunction_unsupported_rdna_generation
(rdna_gen: int)
tilelang/rocm/intrinsics/wmma_layout.py:319
↓ 7 callersFunction_validate_packed_x2_args
Validate that all arguments are PrimExpr with a supported packed x2 dtype.
tilelang/language/math_intrinsics.py:339
↓ 7 callersFunction_validate_rounding_mode
Validate that the rounding mode is one of the supported IEEE modes
tilelang/language/math_intrinsics.py:7
↓ 7 callersFunctionassert_normalize_to_iter_sum
Assert the result of arith.normalize_to_iter_sum is correct Parameters ---------- index : tvm.tirx.PrimExpr The index to be norma
testing/python/arith/test_arith_iter_affine_map.py:1171
↓ 7 callersFunctionast_get_span
(ast: ast.AST)
tilelang/language/eager/ast.py:23
↓ 7 callersFunctioncheck_cutedsl_available
Fail fast if the CuTeDSL backend cannot be used in this Python environment. Policy: - If the public distribution `nvidia-cutlass-dsl` is inst
tilelang/jit/adapter/cutedsl/checks.py:37
↓ 7 callersMethodemit_config
(self, topk: int)
tilelang/carver/roller/policy/default.py:72
↓ 7 callersMethodenter_frame
(self, frame: AbstractContextManager[Any])
tilelang/language/eager/builder.py:278
↓ 7 callersFunctionflashattn
( batch, heads, seq_q, seq_kv, dim, window_size=None, # None for full attention s
examples/attention_sink/example_mha_sink_fwd_bhsd.py:25
↓ 7 callersMethodfrom_database
Alternative constructor to create a TorchFunction directly from a database.
tilelang/jit/kernel.py:145
↓ 7 callersMethodget_dtype
(self, id=0)
tilelang/carver/roller/node.py:144
↓ 7 callersFunctionget_ldmatrix_offset
( matrix: Literal["A", "B"], row_idx, col_idx, stride, dtype: Literal["float16", "int8", "
tilelang/cuda/intrinsics/layout/utils.py:20
↓ 7 callersMethodget_output_buffers
(self, block: SBlockRV)
tilelang/carver/roller/node.py:71
↓ 7 callersMethodget_output_shape
Get the output shape of the layout. Returns ------- List[int] The shape of the output layout.
tilelang/layout/layout.py:68
↓ 7 callersFunctionget_prim_func_name
Extract a human‑readable function name from a TVM PrimFunc. Prefer the `global_symbol` attribute set on the PrimFunc. If it is missing (
tilelang/utils/language.py:561
↓ 7 callersMethodinverse
Compute the inverse of the current layout transformation. Returns ------- Layout A new Layout object rep
tilelang/layout/layout.py:213
↓ 7 callersMethodis_cache_enabled
(self)
tilelang/env.py:355
↓ 7 callersMethodldmatrix_b
(self, B_local_buf: Buffer, B_shared_buf: Buffer | BufferRegion, ki: PrimExpr, rk: PrimExpr | None = 0)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:381
↓ 7 callersFunctionmake_swizzle_layout
(shared_buf)
examples/gemm/example_gemm_intrinsics.py:11
↓ 7 callersMethodmfma
(self, A_local_buf: Buffer, B_local_buf: Buffer, C_local_buf: Buffer, k_inner: PrimExpr | None = 0)
tilelang/rocm/intrinsics/mfma_macro_generator.py:469
↓ 7 callersMethodmma
(self, A_local_buf: Buffer, B_local_buf: Buffer, C_local_buf: Buffer, k_inner: PrimExpr | None = 0)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:501
↓ 7 callersMethodpop
Pop and return the top item from the stack. Returns: The item at the top of the stack. Raises: Inde
tilelang/autotuner/capture.py:29
↓ 7 callersFunctionraise_mismatch_error
(attribute_name: str, actual_value, expected_value)
tilelang/utils/tensor.py:143
↓ 7 callersFunctionref_program
(A)
testing/python/language/test_tilelang_language_atomic.py:359
↓ 7 callersFunctionrun_ieee_math_test
Test IEEE-compliant math operations with specified rounding modes.
testing/python/math/test_math_ieee_math.py:10
↓ 7 callersMethodstmatrix
(self, C_local_buf, C_buf, pid_m=None, pid_n=None)
tilelang/rocm/intrinsics/wmma_macro_generator.py:333
↓ 7 callersMethodtorch_dtype
Converts the TVM DataType to PyTorch dtype. This method is used when creating PyTorch tensors from KernelParam, as PyTorch's
tilelang/engine/param.py:127
↓ 6 callersMethodCurrent
Get the current (topmost) let frame. Returns: LetFrame: The current let frame Raises: IndexError: If there a
tilelang/language/frame.py:173
↓ 6 callersFunction_auto_broadcast
(a, b, op)
tilelang/language/parser/operation.py:58
↓ 6 callersFunction_candidate_cu_dirs
(nvidia_dir: pathlib.Path)
cmake/find_pip_cuda.py:44
↓ 6 callersFunction_collect_calls
(stmt, op_name: str)
testing/python/transform/test_tilelang_transform_lower_shared_tmem.py:18
↓ 6 callersMethod_get_dtype_abbrv
(self, dtype: str)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:144
↓ 6 callersFunction_get_type_str
(dtype: str)
tilelang/language/parser/operation.py:52
↓ 6 callersFunction_is_one
(x: PrimExpr)
tilelang/carver/matmul_analysis.py:41
↓ 6 callersFunction_make_nan_reduce_kernel
(reduce_fn, M, N, dtype, threads, *, nan_propagate)
testing/python/language/test_tilelang_language_reduce.py:351
↓ 6 callersMethod_post_init
Override base class _post_init to register cleanup via weakref.finalize.
tilelang/jit/adapter/cutedsl/adapter.py:431
↓ 6 callersMethod_pythonic_expr
Convert TVM expression to Python string.
tilelang/jit/adapter/cutedsl/wrapper.py:819
↓ 6 callersFunction_stmatrix
Internal helper for stmatrix operations
tilelang/contrib/cutedsl/ldsm.py:51
↓ 6 callersFunction_substitute_bind_env
Apply the current flat-Bind environment to a statement or expression.
tilelang/transform/decouple_type_cast.py:212
↓ 6 callersFunctionabs
Get absolute value of the input element-wise. Parameters ---------- x : PrimExpr Input argument. span : Optional[Span]
tilelang/language/tir/op.py:2689
↓ 6 callersFunctionadd
(path: str | None)
tilelang/contrib/msvc.py:50
↓ 6 callersFunctionassert_simdgroup_store_correctness
(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32, atol=1e-2)
testing/python/metal/test_metal_simdgroup_store.py:40
↓ 6 callersMethodassign_compile_flags
(self, compile_flags: list[str] | None = None)
tilelang/jit/adapter/libgen.py:44
↓ 6 callersMethodbackward
(ctx, do)
examples/attention_sink/example_mha_sink_bwd_bhsd.py:352
↓ 6 callersMethodbackward
(ctx, dy)
examples/norm/layernorm.py:152
↓ 6 callersMethodbind
(self, name, value, annot=BaseBuilder.empty)
tilelang/language/eager/builder.py:416
↓ 6 callersFunctioncheck_correctness_and_bench
(kernel, N, K, do_bench=True)
examples/gemv/example_gemv.py:320
↓ 6 callersFunctionconst
Declare constexpr variables for dynamic tensor dimensions (eager mode only). In eager mode, use T.const() to declare shape dimensions that w
tilelang/language/eager/builder.py:922
↓ 6 callersMethodeval
(self, val: Any)
tilelang/language/eager/ast.py:186
↓ 6 callersFunctionget_all_factors
(n: int)
tilelang/carver/roller/policy/common.py:4
↓ 6 callersMethodget_block_binding
Returns the block binding for the given dimension. dim=0 corresponds to blockIdx.x, dim=1 to blockIdx.y, and dim=2 to blockIdx.z.
tilelang/language/kernel.py:232
↓ 6 callersMethodget_kernel_source
Returns the source code of the compiled kernel function. Returns ------- str The source code of the comp
tilelang/jit/kernel.py:447
↓ 6 callersMethodget_output_embeddings
(self)
examples/bitnet-1.58b/modeling_bitnet.py:1148
↓ 6 callersMethodget_rstep
(self, node)
tilelang/carver/roller/hint.py:100
↓ 6 callersMethodget_tcgen5_mma_meta
Query the FFI for TCGEN5MMA atom metadata (atom_m, atom_n, atom_k, enable_ws, enable_2cta), and record them in `self.meta`.
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:572
↓ 6 callersMethodis_float4_e2m1_unpacked
(self)
tilelang/language/dtypes.py:21
↓ 6 callersFunctionis_hip_target
(target: Target)
tilelang/jit/adapter/utils.py:101
↓ 6 callersMethodis_none
(self)
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:56
↓ 6 callersFunctionkernel
( logits: T.Tensor((n_tok, n_exp), T.float32), out_max: T.Tensor((n_tok,), T.float32),
testing/python/amd/test_tilelang_hip_codegen.py:58
↓ 6 callersFunctionkernel_fn
()
testing/python/transform/test_tilelang_transform_decouple_type_cast.py:237
↓ 6 callersMethodldmatrix_a
(self, A_local_buf: Buffer, A_shared_buf: Buffer | BufferRegion, ki: PrimExpr, rk: PrimExpr | None = 0)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:269
↓ 6 callersFunctionlegalize
(arg)
tilelang/language/gemm_op.py:344
↓ 6 callersMethodmake_mma_load_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:670
← previousnext →201–300 of 7,254, ranked by callers