import pytest
import asc
import asc.runtime.config as config
import asc.lib.runtime as rt
try:
import torch
except ModuleNotFoundError:
pytest.skip("torch is not installed", allow_module_level=True)
@asc.jit
def transdata_to_5hd_kernel(src: asc.GlobalAddress, dst: asc.GlobalAddress,
n: asc.ConstExpr[int], c: asc.ConstExpr[int],
h: asc.ConstExpr[int], w: asc.ConstExpr[int],
c0: asc.ConstExpr[int]):
data_size = n * c * h * w
src_gm = asc.GlobalTensor()
dst_gm = asc.GlobalTensor()
src_gm.set_global_buffer(src)
dst_gm.set_global_buffer(dst)
pipe = asc.TPipe()
in_queue_src = asc.TQue(asc.TPosition.VECIN, 1)
out_queue_dst = asc.TQue(asc.TPosition.VECOUT, 1)
work_queue_src1 = asc.TQue(asc.TPosition.VECCALC, 1)
work_queue_src2 = asc.TQue(asc.TPosition.VECCALC, 1)
pipe.init_buffer(que=in_queue_src, num=1, len=data_size * src.dtype.sizeof())
pipe.init_buffer(que=out_queue_dst, num=1, len=data_size * dst.dtype.sizeof())
pipe.init_buffer(que=work_queue_src1, num=1, len=16 * asc.uint64.sizeof())
pipe.init_buffer(que=work_queue_src2, num=1, len=16 * asc.uint64.sizeof())
copy_in(src_gm, in_queue_src, data_size)
compute(dst_gm, in_queue_src, out_queue_dst, work_queue_src1, work_queue_src2, n, c, h, w, c0)
copy_out(dst_gm, out_queue_dst, data_size)
@asc.jit
def copy_in(src_gm: asc.GlobalTensor, in_queue_src: asc.TQue, src_data_size: int):
src_local = in_queue_src.alloc_tensor(src_gm.dtype)
asc.data_copy(src_local, src_gm, count=src_data_size)
in_queue_src.enque(src_local)
@asc.jit
def compute(dst_gm: asc.GlobalTensor, in_queue_src: asc.TQue, out_queue_dst: asc.TQue,
work_queue_src1: asc.TQue, work_queue_src2: asc.TQue,
n: int, c: int, h: int, w: int, c0: int):
src_local = in_queue_src.deque(dst_gm.dtype)
dst_local = out_queue_dst.alloc_tensor(dst_gm.dtype)
params = asc.TransDataTo5HDParams(
dst_high_half=False,
src_high_half=False,
repeat_times=16,
dst_rep_stride=16,
src_rep_stride=1
)
for j in range(4):
dst_addr_local = work_queue_src1.alloc_tensor(asc.uint64)
src_addr_local = work_queue_src2.alloc_tensor(asc.uint64)
for i in range(16):
dst_offset = j * c0 * h * w + w * i
dst_addr = dst_local[dst_offset].get_phy_addr()
dst_addr_local.set_value(i, dst_addr)
for i in range(16):
src_offset = j * c0 * h * w + h * w * i
src_addr = src_local[src_offset].get_phy_addr()
src_addr_local.set_value(i, src_addr)
asc.trans_data_to_5hd(dst_addr_local, src_addr_local, params)
work_queue_src1.free_tensor(dst_addr_local)
work_queue_src2.free_tensor(src_addr_local)
out_queue_dst.enque(dst_local)
in_queue_src.free_tensor(src_local)
@asc.jit
def copy_out(dst_gm: asc.GlobalTensor, out_queue_dst: asc.TQue, dst_data_size: int):
dst_local = out_queue_dst.deque(dst_gm.dtype)
asc.data_copy(dst_gm, dst_local, count=dst_data_size)
out_queue_dst.free_tensor(dst_local)
def transdata_to_5hd_launch(x: torch.Tensor, c0: int = 16) -> torch.Tensor:
n, c, h, w = x.shape
if c % c0 != 0:
raise ValueError(f"Channel dimension {c} must be divisible by c0={c0}")
z = torch.zeros_like(x)
use_core_num = 1
transdata_to_5hd_kernel[use_core_num, rt.current_stream()](
x, z, n, c, h, w, c0
)
return z
param_list = [
torch.float16,
torch.int16,
]
BACKENDS = [
config.Backend.NPU,
]
@pytest.mark.parametrize("dtype", param_list)
@pytest.mark.parametrize("backend", BACKENDS)
def test_transdata_to_5hd_addr_tensor(dtype, backend: config.Backend):
config.set_platform(backend)
device = "npu" if config.Backend(backend) == config.Backend.NPU else "cpu"
n, c, h, w = 2, 32, 16, 16
c0 = 16
c1 = c // c0
if dtype in {torch.float16, torch.float32}:
x_nchw = torch.randn((n, c, h, w), dtype=dtype, device=device)
else:
x_nchw = torch.randint(0, 99, (n, c, h, w), dtype=dtype, device=device)
z_nc1hwc0 = transdata_to_5hd_launch(x_nchw, c0).reshape(n, c1, h, w, c0)
expected = x_nchw.reshape(n, c1, c0, h, w)
expected = expected.permute((0, 1, 3, 4, 2))
assert torch.allclose(z_nc1hwc0, expected)