MCPcopy Create free account

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

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

Methodmaybe_contiguous
(x)
examples/attention_sink/example_gqa_sink_bwd_bhsd.py:335
Methodmaybe_desc
Record descriptor aliases while matching declaration parameters.
tilelang/jit/adapter/cutedsl/wrapper.py:934
Methodmbar
(self)
tilelang/tileop/gemm/gemm_base.py:176
Methodmbarptr
(self)
tilelang/tileop/gemm/gemm_base.py:172
Functionmbarrier_arrive_expect_tx
Arrive at a memory barrier and expect completion of async transactions.
tilelang/language/builtin.py:523
Functionmbarrier_cp_async_arrive
(mbar_ptr: Pointer, *, loc=None, ip=None)
tilelang/contrib/cutedsl/cpasync.py:324
Functionmbarrier_expect_tx
Set expected transaction count for memory barrier. Args: mbarrier: BarrierType The memory barrier to expect transaction count
tilelang/language/builtin.py:507
Functionmbarrier_wait
Waits on a mbarrier with a specified phase (blocking loop). Uses inline PTX to loop until the try_wait succeeds. The CUDA backend does: while
tilelang/contrib/cutedsl/cpasync.py:304
Functionmetadata_16bit_load_32x2_to_shared_16x2_layout_16bit
(thread_id: int, local_id: int)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:96
Functionmetadata_16bit_load_32x2_to_shared_16x4_layout_8bit
(thread_id: int, local_id: int)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:111
Functionmetadata_32bit_load_32x1_to_shared_16x2_layout_8bit
(thread_id: int, local_id: int)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:118
Functionmetadata_8bit_load_32x4_to_shared_16x4_layout_16bit
(thread_id: int, local_id: int)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:89
Functionmetadata_8bit_load_32x4_to_shared_16x4_layout_32bit
(thread_id: int, local_id: int)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:75
Functionmetadata_8bit_load_32x4_to_shared_16x8_layout_8bit
(thread_id: int, local_id: int)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:104
Functionmfma_store_index_map
(thread_id, local_id)
tilelang/rocm/intrinsics/utils.py:10
Functionmfma_store_index_map_32x32
(thread_id, local_id)
tilelang/rocm/intrinsics/utils.py:14
Functionmin2
Packed element-wise minimum.
tilelang/language/math_intrinsics.py:399
Functionmin_value
minimum value of dtype Parameters ---------- dtype : str The data type. span : Optional[Span] The location of this o
tilelang/language/tir/op.py:2139
Functionminimal_kernel
(A: T.Buffer((1,), "float16"))
testing/python/issue/test_tilelang_issue_1810.py:18
Functionmixed_async_num_stages_one
( x: T.Tensor((M, K), T.float16), meta: T.Tensor((2 * cp_async_bytes,), T.uint8), y: T
testing/python/issue/test_tilelang_issue_tma_no_ws.py:252
Functionmixed_gemm_shared_barrier
( A: T.Tensor((M, K), T.float16), B: T.Tensor((K, N), T.float16), C: T.Tensor((M, N),
testing/python/issue/test_tilelang_issue_tma_no_ws.py:311
Methodmma
Issue SM70 MMA operations using the loaded A/B fragments.
tilelang/cuda/intrinsics/macro/mma_sm70_macro_generator.py:280
Methodmma
(self, A_local_buf, B_local_buf, C_local_buf)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:1176
Functionmma_32x8_to_shared_16x16_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_layout.py:130
Functionmma_32x8_to_shared_16x16_layout_fp16
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sm70_layout.py:25
Functionmma_32x8_to_shared_16x16_layout_fp32
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sm70_layout.py:19
Functionmma_fill
TVM intrinsic for zero-initalizing an MMA accumulation register Parameters ---------- dtype : str The data type of the result.
tilelang/language/tir/op.py:1445
Functionmma_load_a_32x4_to_shared_16x4_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sm70_layout.py:31
Functionmma_load_b_32x16_to_shared_16x32_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_layout.py:170
Functionmma_load_b_32x4_to_shared_16x4_layout_trans
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sm70_layout.py:37
Functionmma_load_b_32x4_to_shared_16x8_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_layout.py:142
Functionmma_load_b_32x4_to_shared_4x16_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sm70_layout.py:43
Functionmma_load_b_32x8_to_shared_16x16_layout
groupID = %laneid >> 2 threadID_in_group = %laneid % 4 row = (threadID_in_group * 2) + (i & 0x1) for bi where i <
tilelang/cuda/intrinsics/layout/mma_layout.py:176
Methodmma_load_layout
(i, j)
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:320
Methodmma_load_layout
(i, j)
tilelang/cuda/intrinsics/macro/mma_sp_macro_generator.py:238
Methodmma_load_layout
(i, j)
tilelang/cuda/intrinsics/macro/wgmma_sp_macro_generator.py:443
Methodmma_num_inst_m
Number of MMA instruction atoms along the M dimension.
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:515
Methodmma_num_inst_n
Number of MMA instruction atoms along the N dimension.
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:520
Functionmma_op
( a_ptr: Pointer, a_offset, b_ptr: Pointer, b_offset, c_ptr: Pointer,
tilelang/contrib/cutedsl/ptx_mma.py:91
Functionmma_sp_load_a_32x16_to_shared_16x64_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:49
Functionmma_sp_load_a_32x4_to_shared_16x16_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:41
Functionmma_sp_load_a_32x8_to_shared_16x32_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:45
Functionmma_sp_load_b_32x16_to_shared_16x32_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:59
Functionmma_sp_load_b_32x32_to_shared_16x64_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:65
Functionmma_sp_load_b_32x8_to_shared_16x16_layout
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/mma_sp_layout.py:53
Functionmma_sp_op
( a_ptr: Pointer, a_offset, b_ptr: Pointer, b_offset, c_ptr: Pointer,
tilelang/contrib/cutedsl/ptx_mma.py:180
Functionmma_store
TVM intrinsic for storing the result of PTX MMA into a destination pointer Parameters ---------- dtype : str The data type of the
tilelang/language/tir/op.py:1411
Functionmma_store_index_map_fp64
(thread_id, local_id)
tilelang/cuda/intrinsics/layout/utils.py:86
Functionmoe_forward_tilelang_routed
( input, routed_expert_gate, routed_expert_up, routed_expert_down, routed_expert_weights,
examples/fusedmoe/example_fusedmoe_tilelang.py:92
Functionmoe_forward_tilelang_shared
( input, shared_W_gate, shared_W_up, shared_W_down, up_logits, output, dtype,
examples/fusedmoe/example_fusedmoe_tilelang.py:12
Functionmul2
Packed element-wise multiply (x * y).
tilelang/language/math_intrinsics.py:374
Functionmxfp8_blockscaled_gemm
1D-1D Block-scaled MXFP8 GEMM. A: [M, K] in FP8 (E4M3 or E5M2) B: [K, N] in FP8 (E4M3 or E5M2), or [N, K] when transpose_B=True SFA:
examples/blockscaled_gemm_sm100/gemm_mxfp8_blockscaled_1d1d.py:13
Functionmxfp8_blockscaled_gemm_2cta
( A, B, SFA, SFB, block_M, block_N, block_K, in_dtype, out_dtype, accu
examples/blockscaled_gemm_sm100/gemm_mxfp8_blockscaled_1d1d.py:170
Functionmxfp8_blockscaled_gemm_2cta_persistent
( A, B, SFA, SFB, block_M, block_N, block_K, in_dtype, out_dtype, accu
examples/blockscaled_gemm_sm100/gemm_mxfp8_blockscaled_1d1d.py:319
Methodn_max_threads
(self)
tilelang/jit/kernel.py:670
Methodn_regs
(self)
tilelang/jit/kernel.py:660
Methodn_spills
(self)
tilelang/jit/kernel.py:665
Functionnaive_gemv
( A, B, BLOCK_N: int, BLOCK_K: int, dtype: T.dtype = T.float16, accum_dtype: T.dtype =
examples/gemv/example_gemv.py:15
Functionnaive_nsa_simple
r""" Args: q (torch.Tensor): queries of shape `[B, T, HQ, K]` if `head_first=False` else `[B, HQ, T, K]`. k (torch.Ten
examples/deepseek_nsa/reference.py:152
Functionnaive_splitk_gemv
( A, B, BLOCK_N: int, BLOCK_K: int, dtype: T.dtype = T.float16, accum_dtype: T.dtype =
examples/gemv/example_gemv.py:46
Functionnamed_barrier_arrive
CTA named barrier one-sided arrive (bar.arrive). Signals that the calling threads have arrived at the named barrier without waiting for other
tilelang/language/builtin.py:964
Methodnative_forward
(self, input)
examples/bitnet-1.58b/utils_quant.py:173
Functionnearbyint
Round elements of the array to the nearest integer. This intrinsic uses llvm.nearbyint instead of llvm.round which is faster but will results
tilelang/language/tir/op.py:2830
Functionnegative_index_before
(A: T.Buffer((16,), T.float32), B: T.Buffer((16,), T.float32))
testing/python/language/test_tilelang_language_negative_index.py:8
Functionnegative_index_expected
(A: T.Buffer((16,), T.float32), B: T.Buffer((16,), T.float32))
testing/python/language/test_tilelang_language_negative_index.py:14
Functionnegative_index_loop_before
(A: T.Buffer((16,), T.float32), B: T.Buffer((4,), T.float32))
testing/python/language/test_tilelang_language_negative_index.py:20
Functionnegative_index_loop_expected
(A: T.Buffer((16,), T.float32), B: T.Buffer((4,), T.float32))
testing/python/language/test_tilelang_language_negative_index.py:27
Functionnegative_index_symbolic_before
(shift: T.int32, A: T.Buffer((16,), T.float32), B: T.Buffer((16,), T.float32))
testing/python/language/test_tilelang_language_negative_index.py:34
Functionnext_power_of_2
(x: int)
tilelang/math/__init__.py:1
Functionnextafter
Return the next floating-point value after x1 towards x2. Parameters ---------- x1 : PrimExpr Input argument. x2 : PrimExpr
tilelang/language/tir/op.py:2856
Functionnum_gpus_available
Get number of GPUs without initializing the CUDA context in current process.
examples/bitnet-1.58b/vllm_workspace/conftest.py:583
Methodnum_outputs
(self)
tilelang/carver/roller/node.py:167
Methodnum_threads
Returns the total number of threads.
tilelang/language/kernel.py:260
Methodoffset_A
(self)
tilelang/tileop/gemm/gemm_base.py:144
Methodoffset_A
(self)
tilelang/tileop/gemm/__init__.py:94
Methodoffset_A
(self)
tilelang/tileop/gemm_sp/gemm_sp_base.py:116
Methodoffset_B
(self)
tilelang/tileop/gemm/gemm_base.py:148
Methodoffset_B
(self)
tilelang/tileop/gemm/__init__.py:98
Methodoffset_B
(self)
tilelang/tileop/gemm_sp/gemm_sp_base.py:120
Methodopaque
The opaque block axis defining function. Parameters ---------- dom : Union[Range, List[PrimExpr], Tuple[PrimExpr]]
tilelang/language/ast/ir.py:620
Methodout_idx
(self)
tilelang/jit/kernel.py:618
Methodoutput_nodes
Returns the output nodes associated with this template. Returns: List[OutputNode]: The output nodes.
tilelang/carver/template/base.py:176
Methodoutputs
(self)
tilelang/carver/roller/node.py:133
Methodoverride
(self, name: str)
testing/python/language/test_tilelang_language_eager_ast_signature.py:5
Methodoverride
(self, name: str)
tilelang/language/eager/ast.py:244
Methodoverride
(self, name: str)
tilelang/language/eager/builder.py:747
Functionpack_half2
Pack two half-precision (fp16) values into a single 32-bit value. Corresponds to CUDA's __pack_half2 intrinsic. This packs two fp16 valu
tilelang/contrib/cutedsl/utils.py:138
Methodpanel_width
(self)
tilelang/carver/roller/rasterization.py:14
Functionparallel
The parallel For statement. Parameters ---------- start : PrimExpr The minimum value of iteration. stop : PrimExpr T
tilelang/language/ast/ir.py:704
Functionparallel
The parallel For statement. Parameters ---------- start : PrimExpr The minimum value of iteration. stop : PrimExpr T
tilelang/language/tir/ir.py:32
Functionparallel_nsa
r""" Args: q (torch.Tensor): queries of shape `[B, T, HQ, K]` if `head_first=False` else `[B, HQ, T, K]`. k (torch.Ten
examples/deepseek_nsa/benchmark/benchmark_nsa_fwd.py:229
Functionparallel_nsa_bwd_kernel_dkv
( q, k, v, lse_slc, lse_swa, delta_slc, delta_swa, do_slc, do_swa, dk,
examples/deepseek_nsa/example_triton_nsa_bwd.py:229
Functionparallel_nsa_bwd_kernel_dq
( q, k, v, lse_slc, delta_slc, do_slc, lse_swa, delta_swa, do_swa, dq,
examples/deepseek_nsa/example_triton_nsa_bwd.py:380
Functionparallel_nsa_bwd_kernel_preprocess
(o, do, delta, B: tl.constexpr, V: tl.constexpr)
examples/deepseek_nsa/example_triton_nsa_bwd.py:656
Functionparallel_nsa_fwd_kernel
( q, k, v, o_slc, o_swa, lse_slc, lse_swa, scale, block_indices, block
examples/deepseek_nsa/example_triton_nsa_fwd_varlen.py:32
Functionparallel_nsa_fwd_kernel
( q, k, v, o_slc, o_swa, lse_slc, lse_swa, scale, block_indices, block
examples/deepseek_nsa/example_triton_nsa_bwd.py:32
Functionparallel_nsa_fwd_kernel
( q, k, v, o_slc, o_swa, lse_slc, lse_swa, scale, block_indices, block
examples/deepseek_nsa/example_triton_nsa_fwd.py:32
Functionparallel_nsa_fwd_kernel
( q, k, v, o_slc, o_swa, lse_slc, lse_swa, scale, block_indices, block
examples/deepseek_nsa/benchmark/benchmark_nsa_fwd.py:28
Functionparallel_nsa_kernel_mask
( block_indices, block_counts, block_mask, T: tl.constexpr, H: tl.constexpr, S: tl.con
examples/deepseek_nsa/example_triton_nsa_bwd.py:345
Methodparams
(self)
tilelang/jit/kernel.py:622
← previousnext →5,201–5,300 of 7,254, ranked by callers