tilelang.carver.arch.cuda¶

Attributes¶

volta_tensorcore_supported

ampere_tensorcore_supported

ada_tensorcore_supported

hopper_tensorcore_supported

Classes¶

TensorInstruction

CUDA

Represents the architecture of a computing device, capturing various hardware specifications.

Functions¶

check_sm_version(arch)

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.check_sm_version(arch)¶
Parameters:

arch (str)

Return type:

int

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.volta_tensorcore_supported = [('float16', 'float32'), ('float16', 'float16')]¶
tilelang.carver.arch.cuda.ampere_tensorcore_supported = [('bfloat16', 'float32'), ('float16', 'float32'), ('float16', 'float16'), ('int8', 'int32'),...¶
tilelang.carver.arch.cuda.ada_tensorcore_supported = [('bfloat16', 'float32'), ('float16', 'float32'), ('float16', 'float16'), ('int8', 'int32'),...¶
tilelang.carver.arch.cuda.hopper_tensorcore_supported = [('bfloat16', 'float32'), ('float16', 'float32'), ('float16', 'float16'), ('int8', 'int32'),...¶
tilelang.carver.arch.cuda.is_tensorcore_supported_precision(in_dtype, accum_dtype, arch)¶
Parameters:
Return type:

bool

class tilelang.carver.arch.cuda.TensorInstruction(name, shape)¶

Bases: object

Parameters:

  • name (str)

  • shape (List[int])

name: str¶
shape: List[int]¶
class tilelang.carver.arch.cuda.CUDA(target)¶

Bases: tilelang.carver.arch.arch_base.TileDevice

Represents the architecture of a computing device, capturing various hardware specifications.

Parameters:

target (Union[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__()¶