tilelang.intrinsics.mma_sp_layout

函式

shared_16x16_to_mma_sp_layout_sr_a(i, j)
shared_16x16_to_mma_sp_layout_sr_b(i, j)
shared_16x32_to_mma_sp_layout_sr_a(i, j)
shared_16x32_to_mma_sp_layout_sr_b(i, j)
shared_16x64_to_mma_sp_layout_sr_a(i, j)
shared_16x64_to_mma_sp_layout_sr_b(i, j)
mma_sp_load_a_32x4_to_shared_16x16_layout(thread_id, ...)
mma_sp_load_a_32x8_to_shared_16x32_layout(thread_id, ...)
mma_sp_load_a_32x16_to_shared_16x64_layout(thread_id, ...)
mma_sp_load_b_32x8_to_shared_16x16_layout(thread_id, ...)
mma_sp_load_b_32x16_to_shared_16x32_layout(thread_id, ...)
mma_sp_load_b_32x32_to_shared_16x64_layout(thread_id, ...)
get_logical_id_32bit(thread_id)
metadata_8bit_load_32x4_to_shared_16x4_layout_32bit(...)
metadata_16bit_load_32x2_to_shared_16x2_layout_32bit(...)
metadata_8bit_load_32x4_to_shared_16x4_layout_16bit(...)
metadata_16bit_load_32x2_to_shared_16x2_layout_16bit(...)
get_logical_id_8bit(thread_id)
metadata_8bit_load_32x4_to_shared_16x4_layout_8bit(...)
metadata_16bit_load_32x2_to_shared_16x4_layout_8bit(...)
metadata_32bit_load_32x1_to_shared_16x2_layout_8bit(...)
ldmatrix_trans_32x8_to_shared_16x16_layout(thread_id, ...)
ldmatrix_32x16_to_shared_32x16_layout(thread_id, local_id)
ldmatrix_trans_32x16_to_shared_16x32_layout(thread_id, ...)
ldmatrix_trans_32x32_to_shared_shared_16x64_layout(...)
get_ldmatrix_offset_b(matrix, row_idx, col_idx, stride)

Module Contents

tilelang.intrinsics.mma_sp_layout.shared_16x16_to_mma_sp_layout_sr_a(i, j)

tilelang.intrinsics.mma_sp_layout.shared_16x16_to_mma_sp_layout_sr_b(i, j)

tilelang.intrinsics.mma_sp_layout.shared_16x32_to_mma_sp_layout_sr_a(i, j)

tilelang.intrinsics.mma_sp_layout.shared_16x32_to_mma_sp_layout_sr_b(i, j)

tilelang.intrinsics.mma_sp_layout.shared_16x64_to_mma_sp_layout_sr_a(i, j)

tilelang.intrinsics.mma_sp_layout.shared_16x64_to_mma_sp_layout_sr_b(i, j)

tilelang.intrinsics.mma_sp_layout.mma_sp_load_a_32x4_to_shared_16x16_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.mma_sp_load_a_32x8_to_shared_16x32_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.mma_sp_load_a_32x16_to_shared_16x64_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.mma_sp_load_b_32x8_to_shared_16x16_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.mma_sp_load_b_32x16_to_shared_16x32_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.mma_sp_load_b_32x32_to_shared_16x64_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.get_logical_id_32bit(thread_id)

參數:

thread_id (int)

回傳型別:

int

tilelang.intrinsics.mma_sp_layout.metadata_8bit_load_32x4_to_shared_16x4_layout_32bit(thread_id, local_id)

參數:

• thread_id (int)

• local_id (int)

回傳型別:

tuple[int, int]

tilelang.intrinsics.mma_sp_layout.metadata_16bit_load_32x2_to_shared_16x2_layout_32bit(thread_id, local_id)

參數:

• thread_id (int)

• local_id (int)

回傳型別:

tuple[int, int]

tilelang.intrinsics.mma_sp_layout.metadata_8bit_load_32x4_to_shared_16x4_layout_16bit(thread_id, local_id)

參數:

• thread_id (int)

• local_id (int)

回傳型別:

tuple[int, int]

tilelang.intrinsics.mma_sp_layout.metadata_16bit_load_32x2_to_shared_16x2_layout_16bit(thread_id, local_id)

參數:

• thread_id (int)

• local_id (int)

回傳型別:

tuple[int, int]

tilelang.intrinsics.mma_sp_layout.get_logical_id_8bit(thread_id)

參數:

thread_id (int)

回傳型別:

int

tilelang.intrinsics.mma_sp_layout.metadata_8bit_load_32x4_to_shared_16x4_layout_8bit(thread_id, local_id)

參數:

• thread_id (int)

• local_id (int)

回傳型別:

tuple[int, int]

tilelang.intrinsics.mma_sp_layout.metadata_16bit_load_32x2_to_shared_16x4_layout_8bit(thread_id, local_id)

參數:

• thread_id (int)

• local_id (int)

回傳型別:

tuple[int, int]

tilelang.intrinsics.mma_sp_layout.metadata_32bit_load_32x1_to_shared_16x2_layout_8bit(thread_id, local_id)

參數:

• thread_id (int)

• local_id (int)

回傳型別:

tuple[int, int]

tilelang.intrinsics.mma_sp_layout.ldmatrix_trans_32x8_to_shared_16x16_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.ldmatrix_32x16_to_shared_32x16_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.ldmatrix_trans_32x16_to_shared_16x32_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.ldmatrix_trans_32x32_to_shared_shared_16x64_layout(thread_id, local_id)

tilelang.intrinsics.mma_sp_layout.get_ldmatrix_offset_b(matrix, row_idx, col_idx, stride, dtype='float16', transposed=False)

參數:

• matrix (Literal['B'])

• dtype (Literal['float16', 'int8'])

• transposed (bool)