tilelang.contrib.cutedsl

Submodules

• tilelang.contrib.cutedsl.cpasync
• tilelang.contrib.cutedsl.gemm_V1
• tilelang.contrib.cutedsl.ldsm
• tilelang.contrib.cutedsl.math
• tilelang.contrib.cutedsl.mbar
• tilelang.contrib.cutedsl.reduce
• tilelang.contrib.cutedsl.threadblock_swizzle

Attributes

BYTES_PER_TENSORMAP
BYTES_PER_POINTER

Functions

make_filled_tensor(shape, value)
make_tensor_at_offset(ptr, offset, shape[, div_by])
shuffle_elect(thread_extent)
sync_thread_partial([barrier_id, thread_count])
pack_half2(x, y)	Pack two half-precision (fp16) values into a single 32-bit value.
AtomicAdd(ptr, value, *[, loc, ip])

Package Contents

tilelang.contrib.cutedsl.BYTES_PER_TENSORMAP = 128

tilelang.contrib.cutedsl.BYTES_PER_POINTER = 8

tilelang.contrib.cutedsl.make_filled_tensor(shape, value)

tilelang.contrib.cutedsl.make_tensor_at_offset(ptr, offset, shape, div_by=1)

Parameters:

ptr (threadblock_swizzle.cute.Pointer)

tilelang.contrib.cutedsl.shuffle_elect(thread_extent)

tilelang.contrib.cutedsl.sync_thread_partial(barrier_id=None, thread_count=None)

tilelang.contrib.cutedsl.pack_half2(x, y)

Pack two half-precision (fp16) values into a single 32-bit value. Corresponds to CUDA’s __pack_half2 intrinsic.

This packs two fp16 values into a single int32 by treating the fp16 bits as raw data and concatenating them.

tilelang.contrib.cutedsl.AtomicAdd(ptr, value, *, loc=None, ip=None)

Parameters:

• ptr (threadblock_swizzle.cute.Pointer)

• value (math.Numeric)