tilelang.carver.arch.cuda

Classes

CUDA

Functions

is_cuda_arch(arch)
is_volta_arch(arch)
is_ampere_arch(arch)
is_ada_arch(arch)
is_hopper_arch(arch)
has_mma_support(arch)
is_tensorcore_supported_precision(in_dtype, ...)

Module Contents

tilelang.carver.arch.cuda.is_cuda_arch(arch)

Parameters:

arch (tilelang.carver.arch.arch_base.TileDevice)

Return type:

bool

tilelang.carver.arch.cuda.is_volta_arch(arch)

Parameters:

arch (tilelang.carver.arch.arch_base.TileDevice)

Return type:

bool

tilelang.carver.arch.cuda.is_ampere_arch(arch)

Parameters:

arch (tilelang.carver.arch.arch_base.TileDevice)

Return type:

bool

tilelang.carver.arch.cuda.is_ada_arch(arch)

Parameters:

arch (tilelang.carver.arch.arch_base.TileDevice)

Return type:

bool

tilelang.carver.arch.cuda.is_hopper_arch(arch)

Parameters:

arch (tilelang.carver.arch.arch_base.TileDevice)

Return type:

bool

tilelang.carver.arch.cuda.has_mma_support(arch)

Parameters:

arch (tilelang.carver.arch.arch_base.TileDevice)

Return type:

bool

tilelang.carver.arch.cuda.is_tensorcore_supported_precision(in_dtype, accum_dtype, arch)

Parameters:

• in_dtype (str)

• accum_dtype (str)

• arch (tilelang.carver.arch.arch_base.TileDevice)

Return type:

bool

class tilelang.carver.arch.cuda.CUDA(target)

Bases: tilelang.carver.arch.arch_base.TileDevice

Parameters:

target (tvm.target.Target | str)

target

sm_version

name

device: tvm.runtime.Device

platform: str = 'CUDA'

smem_cap

compute_max_core

warp_size

compute_capability

reg_cap: int = 65536

max_smem_usage: int

sm_partition: int = 4

l2_cache_size_bytes: int

transaction_size: list[int] = [32, 128]

bandwidth: list[int] = [750, 12080]

available_tensor_instructions: list[TensorInstruction] = None

get_avaliable_tensorintrin_shapes()

__repr__()