MCPcopy Create free account

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

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

↓ 2 callersMethod_make_emitter
(self, target: Target, thread_nums: int, thread_var=None)
tilelang/rocm/op/gemm/gemm_wmma.py:23
↓ 2 callersFunction_make_finalize_reducer_kernel
(block_M, block_N, dtype, op, batch)
testing/python/language/test_tilelang_language_reduce.py:208
↓ 2 callersFunction_make_frontend_cache_key
( frontend_key_data: dict, *, target: Target, target_host: str | Target | None, execution_
tilelang/cache/__init__.py:77
↓ 2 callersMethod_make_mma_emitter
(self, target: Target, thread_nums: int, thread_var: tirx.Var | None = None)
tilelang/cuda/op/gemm/gemm_mma.py:23
↓ 2 callersFunction_make_no_overlap_kernel
(disable_reuse: bool)
testing/python/transform/test_tilelang_transform_disable_memory_reuse.py:51
↓ 2 callersFunction_make_rs_kernel
(M: int, num_threads: int, target_dtype: str)
testing/python/language/test_tilelang_cast_rounding.py:134
↓ 2 callersMethod_make_smem_layout_AB
(self, dtype, major_mode, copy_bits, smem_tiler)
tilelang/contrib/cutedsl/gemm_v1.py:259
↓ 2 callersFunction_make_sync_tcgen05_kernel
(gemm_op)
testing/python/language/test_tilelang_language_tcgen05_gemm.py:9
↓ 2 callersFunction_make_ternary_kernel
Build a kernel: D[idx] = op(A[idx], B[idx], C[idx]).
testing/python/cuda/test_cuda_f32x2_intrinsics.py:56
↓ 2 callersMethod_make_tiled_mma
(self, is_rsMode=False)
tilelang/contrib/cutedsl/gemm_v1.py:466
↓ 2 callersFunction_make_unary_kernel
Build a kernel: C[idx] = op(A[idx]).
testing/python/cuda/test_cuda_f32x2_intrinsics.py:74
↓ 2 callersFunction_make_wgmma_kernel
(gemm_op)
testing/python/language/test_tilelang_language_wgmma_gemm.py:9
↓ 2 callersFunction_matmul_kernel
Return a prim_func for pipelined GEMM using T.copy (global->shared).
testing/python/amd/test_tilelang_gfx950_copy_async.py:19
↓ 2 callersFunction_merge_two_bounds
(x: arith.ConstIntBound, y: arith.ConstIntBound)
tilelang/carver/roller/shape_inference/tir.py:43
↓ 2 callersFunction_nd_to_2d
Convert an N-D index to (row, col) in the flattened 2-D view.
tilelang/tools/plot_layout.py:330
↓ 2 callersFunction_normalize_bindings
Return a bare Var when we only have a single binding so that users may write either `with T.Kernel(...) as pid:` or `with T.Kernel(...) as (p
tilelang/language/kernel.py:87
↓ 2 callersFunction_normalize_cluster_dims
( cluster_dims: int | tuple[int, int, int] | list[int] | None, )
tilelang/language/kernel.py:117
↓ 2 callersFunction_normalize_threads
( threads: int | list[int] | tuple | None, *, is_cpu: bool, )
tilelang/language/kernel.py:98
↓ 2 callersFunction_num_non_one
(exts: list)
tilelang/utils/language.py:507
↓ 2 callersFunction_pack_bf16_high
Pack high 16-bits of r0 and r1 into one uint32.
tilelang/contrib/cutedsl/quantize.py:241
↓ 2 callersFunction_pack_bf16_low
Pack low 16-bits of r0 and r1 into one uint32.
tilelang/contrib/cutedsl/quantize.py:260
↓ 2 callersFunction_pack_int4
(tensor: torch.Tensor)
testing/python/kernel/test_tilelang_kernel_gemm_sm75.py:36
↓ 2 callersFunction_packed_fp4_bytes_to_f32
Decode packed fp4 bytes (m, k//2) to logical (m, k) float32.
maint/gemm/correctness_evaluation_tcgen05_fp4_unpacked.py:185
↓ 2 callersFunction_packed_storage_factor
(buffer: Buffer, value)
tilelang/language/eager/builder.py:55
↓ 2 callersFunction_parse_device_list
(raw: str)
maint/scripts/pytest_cuda_scheduler.py:105
↓ 2 callersMethod_parse_phase2_key
(self, **kwargs)
tilelang/language/eager/builder.py:1074
↓ 2 callersFunction_patch_prim_func_attrs
Attach function-level out_idx, pass_configs and compile_flags as PrimFunc attrs.
tilelang/language/eager/builder.py:1011
↓ 2 callersMethod_plan_vectorize
Plans vectorization for a given PrimFuncNode based on the TileDict configuration and block size. Parameters ----------
tilelang/carver/roller/policy/default.py:757
↓ 2 callersFunction_prepare_scan_args
(src: BufferLikeType, dst: BufferLikeType | None, dim: int, op_name: str)
tilelang/language/scan_op.py:63
↓ 2 callersMethod_process_buffer_device
Extract information about buffer devices from the TIR function. Maps buffer variables to their corresponding devices.
tilelang/jit/adapter/cython/adapter.py:312
↓ 2 callersMethod_process_buffer_dtype
Extract information about buffer dtypes from the TIR function. Maps buffer variables to their corresponding dtypes.
tilelang/jit/adapter/cython/adapter.py:245
↓ 2 callersMethod_process_dynamic_symbolic
Extract information about dynamic shapes from the TIR function. Maps symbolic variables to their corresponding (id, buffer_index, dimension,
tilelang/jit/adapter/tvm_ffi.py:118
↓ 2 callersMethod_process_dynamic_symbolic
Extract information about dynamic shapes from the TIR function. Maps symbolic variables to their corresponding (id, buffer_index, dimension,
tilelang/jit/adapter/cython/adapter.py:216
↓ 2 callersMethod_process_dynamic_symbolic
Extract information about dynamic shapes from the TIR function. Maps symbolic variables to their corresponding (buffer_index, shape_dimension
tilelang/jit/adapter/nvrtc/adapter.py:159
↓ 2 callersMethod_process_dynamic_symbolic
Extract information about dynamic symbols from the TIR function. We follow the same ordering semantics as `TLCUDASourceWrapper.get_dynamic_sy
tilelang/jit/adapter/cutedsl/adapter.py:173
↓ 2 callersMethod_process_ptr_map
Extract information about pointer arguments from the TIR function. Maps pointer arguments to their corresponding (buffer_index, shape_dimensi
tilelang/jit/adapter/cython/adapter.py:261
↓ 2 callersMethod_process_static_buffer_infos
Extract information about static shapes from the TIR function. Maps buffer variables to their corresponding static shapes.
tilelang/jit/adapter/cython/adapter.py:275
↓ 2 callersMethod_process_tma_descriptors
Process TMA descriptors and return tensor args and mapping. Returns: Tuple of (tensor_args, tensor_arg_map)
tilelang/jit/adapter/cutedsl/wrapper.py:1411
↓ 2 callersFunction_read_prompts
(filename: str)
examples/bitnet-1.58b/vllm_workspace/conftest.py:36
↓ 2 callersFunction_ref
(A, op)
testing/python/language/test_tilelang_language_reduce.py:22
↓ 2 callersFunction_register_expr_op
(ty: type)
tilelang/language/parser/operation.py:29
↓ 2 callersFunction_remap_buffer
(buf, var_map)
tilelang/metal/transform/metal_fragment_to_simdgroup.py:57
↓ 2 callersFunction_require_cuda_tensor
(shape, dtype=torch.float32)
testing/python/language/test_tilelang_language_parallel.py:42
↓ 2 callersFunction_require_cutedsl
Skip when the CuTeDSL Python stack is unavailable.
testing/python/jit/test_tilelang_jit_cutedsl_host_codegen.py:11
↓ 2 callersFunction_resolve_torch_dtype
(name: str)
benchmark/matmul/benchmark_matmul_sp_compress.py:11
↓ 2 callersFunction_rocm_target_from_arch
(arch: str | None)
tilelang/utils/target.py:82
↓ 2 callersFunction_run_grouped_gemm_ws
(kernel, batch_sizes, K=128, N=128, block_M=64, dtype="float16")
testing/python/transform/test_tilelang_transform_producer_consumer_ws.py:330
↓ 2 callersFunction_run_inject_assumes
(func: tvm.tirx.PrimFunc)
testing/python/transform/test_tilelang_transform_inject_assumes.py:7
↓ 2 callersFunction_run_mma_atom
(M, N, K, in_dtype, out_dtype, accum_dtype)
testing/python/language/test_tilelang_language_atom_mma.py:122
↓ 2 callersFunction_scalar_result_type
(*args)
tilelang/contrib/cutedsl/math.py:41
↓ 2 callersFunction_scan_fragment
( src: BufferLikeType, dst: BufferLikeType, dim: int, reverse: bool, op_key: str, )
tilelang/language/scan_op.py:15
↓ 2 callersMethod_select_gemm_instruction
Select the appropriate GEMM instruction key based on target and thread configuration. The selection logic chooses: 1. TCGEN5MMA for B
tilelang/tileop/gemm/__init__.py:141
↓ 2 callersMethod_select_gemm_instruction
(self, thread_nums: int, target: Target)
tilelang/tileop/gemm_sp/__init__.py:67
↓ 2 callersFunction_select_int64_stress_n
Pick the largest power-of-two element count that fits in free VRAM. Returns ``target_n`` (8 GiB at bf16) when the GPU has enough free memory;
testing/python/language/test_tilelang_language_int64.py:17
↓ 2 callersMethod_select_launcher_args
Select host launcher args needed by the emitted device launches. The Python wrapper still accepts the full PrimFunc signature so callers
tilelang/jit/adapter/cutedsl/wrapper.py:1207
↓ 2 callersFunction_set_worker_process_title
Optionally label worker processes for easier inspection.
maint/scripts/pytest_cuda_scheduler.py:222
↓ 2 callersMethod_should_use_tf32_mma_operand
(dtype: str, accum_dtype: str)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:152
↓ 2 callersMethod_tag_kernel_cache_entry
(kernel: JITKernel, key: str, cache_path: str)
tilelang/cache/kernel_cache.py:430
↓ 2 callersFunction_tcgen05_num_smem_chunks
(smem_src, chunk_elems: int)
tilelang/language/builtin.py:1289
↓ 2 callersFunction_test_kernel
(M, N)
testing/python/issue/test_tilelang_issue_1263.py:5
↓ 2 callersFunction_test_kernel_if_cond
(M, N)
testing/python/issue/test_tilelang_issue_1263.py:27
↓ 2 callersFunction_tflops
(M, N, K, seconds)
benchmark/matmul_metal/benchmark_matmul_metal.py:43
↓ 2 callersFunction_tir_u8_to_i4_to_i8
(nbit: int, val: tirx.PrimExpr, pos: tirx.PrimExpr, dtype: str)
examples/dequantize_gemm/example_dequant_gemm_w4a8.py:10
↓ 2 callersFunction_tl_vs_sparse_flashattn
(batch, heads, seq_len, dim, vertical_size, slash_size)
examples/minference/example_vertical_slash_sparse_attn.py:17
↓ 2 callersFunction_torch_cummax
(chunk, dim, reverse)
testing/python/language/test_tilelang_language_scan.py:8
↓ 2 callersFunction_trial
(func: Callable)
tilelang/carver/common_schedules.py:111
↓ 2 callersFunction_try_eval
(node: ast.expr, nonlocals: dict[str, Any], globals: dict[str, Any])
tilelang/language/eager/ast.py:248
↓ 2 callersFunction_unpack_sf_u32_1d
(packed_sf, mn, sf_k_blocks)
maint/gemm/correctness_evaluation_tcgen05_fp4_unpacked.py:256
↓ 2 callersFunction_unpack_unsigned_int4_to_half
(qB)
examples/dequantize_gemm/example_dequant_gemm_fine_grained.py:11
↓ 2 callersMethod_update_probas
(probas: dict[int, float], task: Task, is_interesting: bool)
tilelang/autodd.py:562
↓ 2 callersFunction_wgmma_num_c_regs
Number of i32 result registers per thread for a WGMMA op. Each i32 register holds ``32 // elem_bits`` packed elements.
tilelang/contrib/cutedsl/gemm_v2.py:108
↓ 2 callersFunction_windows_compile
(output, objects, options, cwd=None, ccache_env=None)
tilelang/contrib/cc.py:406
↓ 2 callersFunction_with_cutedsl_key
(target: Target | str)
tilelang/utils/target.py:183
↓ 2 callersFunctionactivation_quant
(x, num_bits=8)
examples/bitnet-1.58b/utils_quant.py:24
↓ 2 callersMethodactivation_quant
(self, x, num_bits=8)
examples/bitnet-1.58b/utils_quant.py:157
↓ 2 callersFunctionadd_scalar
(N: int = 4096, BLOCK_N: int = 512, threads: int = 128)
testing/python/autotune/test_tilelang_autotune_scalar_inputs.py:11
↓ 2 callersMethodadd_tag
(self, k: str, v: Any = True)
tilelang/carver/roller/node.py:114
↓ 2 callersFunctionalign_up
(x, y)
examples/blockscaled_gemm_sm100/grouped_gemm_mxfp8_blockscaled_1d1d.py:457
↓ 2 callersFunctionalign_up
(x, y)
examples/blockscaled_gemm_sm100/gemm_mxfp8_blockscaled_1d1d.py:547
↓ 2 callersFunctionalloc_shared
Allocate a shared memory buffer for inter-thread communication. Args: shape (tuple): The shape of the buffer to allocate dtype (s
tilelang/language/allocate.py:34
↓ 2 callersFunctionallow_warp_specialized
(pass_ctx: PassContext | None = None, target: Target | None = None)
tilelang/cuda/pipeline.py:27
↓ 2 callersMethodanalysis
Perform a full analysis of the given IR module or PrimFunc. Args: fn: A TVM IRModule or PrimFunc to analyze.
tilelang/tools/Analyzer.py:208
↓ 2 callersFunctionannotate_l2_hit_ratio
Annotate the L2 hit ratio of the buffer.
tilelang/language/annotations.py:53
↓ 2 callersMethodarg
(self, name, value)
tilelang/language/eager/builder.py:741
↓ 2 callersFunctionassert_matmul_correctness
(M, N, K, block_M, block_N, block_K, in_dtype, out_dtype, accum_dtype)
testing/python/kernel/test_tilelang_kernel_int8_gemm_tcgen5.py:41
↓ 2 callersFunctionassert_similar
(x, y, eps=1e-2, name="tensor", assert_=False, print_=True)
examples/flash_decoding/example_gqa_decode.py:391
↓ 2 callersFunctionassert_tl_gemm_correctness
(M, N, K, block_N, in_dtype, out_dtype, accum_dtype)
examples/deepseek_deepgemm/example_deepgemm_fp8_2xAcc.py:141
↓ 2 callersFunctionassert_tl_matmul_correctness
( M, N, K, in_dtype, out_dtype, accum_dtype=T.float32, a_transposed=False, b_t
testing/python/amd/test_tilelang_gemm_mfma_preshuffle.py:270
↓ 2 callersFunctionast_has_span
(ast: ast.AST)
tilelang/language/eager/ast.py:19
↓ 2 callersFunctionatomic_add_program
(K, M, N, block_M, block_N, dtype=T.float32)
testing/python/language/test_tilelang_language_atomic.py:11
↓ 2 callersFunctionattach_rewrites
(tree: ast.AST, rewrites: list[ASTRewrite])
tilelang/autodd.py:481
↓ 2 callersFunctionattr
Create an attribute node. Parameters ---------- node : Any The node to annotate the attribute. attr_key : str Attrib
tilelang/language/ast/ir.py:1050
↓ 2 callersFunctionauto_inline_consumers
( sch: Schedule, block: SBlockRV, )
tilelang/carver/matmul_analysis.py:82
↓ 2 callersMethodbackward
(ctx, do_slc, do_swa)
examples/deepseek_nsa/example_tilelang_nsa_bwd.py:667
↓ 2 callersMethodbackward
(ctx, grad_output)
examples/grouped_gemm/example_grouped_gemm_bwd.py:91
↓ 2 callersFunctionbasic_shape_kernel
(x)
testing/python/language/test_tilelang_language_subtype.py:18
↓ 2 callersFunctionbenchmark_function
(func, *args, warmup=10, repeat=100)
examples/amd/example_amd_flash_attn_bwd.py:489
↓ 2 callersFunctionbenchmark_nsa
Benchmark the TileLang Sparse Attention implementation.
examples/deepseek_nsa/benchmark/benchmark_nsa_fwd.py:580
↓ 2 callersFunctionbenchmark_triton_nsa
Benchmark the Triton-based TileLang Sparse Attention implementation.
examples/deepseek_nsa/benchmark/benchmark_nsa_fwd.py:681
↓ 2 callersMethodbind_immutable
Bind an immutable tilelang objects. The immutability means the result is usually not changed or re-assigned in a python block.
tilelang/language/eager/builder.py:522
← previousnext →901–1,000 of 7,254, ranked by callers