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(kernel_type=config.KernelType.AIV_ONLY)
def vadd_kernel(x: asc.GlobalAddress, y: asc.GlobalAddress, z: asc.GlobalAddress, block_length: asc.ConstExpr[int],
buffer_num: asc.ConstExpr[int], tile_length: asc.ConstExpr[int], tile_num: asc.ConstExpr[int]):
offset = asc.get_block_idx() * block_length
x_gm = asc.GlobalTensor()
y_gm = asc.GlobalTensor()
z_gm = asc.GlobalTensor()
x_gm.set_global_buffer(x + offset)
y_gm.set_global_buffer(y + offset)
z_gm.set_global_buffer(z + offset)
pipe = asc.TPipe()
in_queue_x = asc.TQue(asc.TPosition.VECIN, buffer_num)
in_queue_y = asc.TQue(asc.TPosition.VECIN, buffer_num)
out_queue_z = asc.TQue(asc.TPosition.VECOUT, buffer_num)
pipe.init_buffer(que=in_queue_x, num=buffer_num, len=tile_length * x.dtype.sizeof())
pipe.init_buffer(que=in_queue_y, num=buffer_num, len=tile_length * y.dtype.sizeof())
pipe.init_buffer(que=out_queue_z, num=buffer_num, len=tile_length * z.dtype.sizeof())
for i in range(tile_num):
copy_in(i, x_gm, y_gm, in_queue_x, in_queue_y, tile_length)
compute(z_gm, in_queue_x, in_queue_y, out_queue_z, tile_length)
copy_out(i, z_gm, out_queue_z, tile_length)
@asc.jit
def copy_in(i: int, x_gm: asc.GlobalAddress, y_gm: asc.GlobalAddress, in_queue_x: asc.TQue, in_queue_y: asc.TQue,
tile_length: asc.ConstExpr[int]):
x_local = in_queue_x.alloc_tensor(x_gm.dtype)
y_local = in_queue_y.alloc_tensor(y_gm.dtype)
asc.data_copy(x_local, x_gm[i * tile_length:], count=tile_length)
asc.data_copy(y_local, y_gm[i * tile_length:], count=tile_length)
in_queue_x.enque(x_local)
in_queue_y.enque(y_local)
@asc.jit
def compute(z_gm: asc.GlobalTensor, in_queue_x: asc.TQue, in_queue_y: asc.TQue, out_queue_z: asc.TQue,
tile_length: asc.ConstExpr[int]):
x_local = in_queue_x.deque(z_gm.dtype)
y_local = in_queue_y.deque(z_gm.dtype)
z_local = out_queue_z.alloc_tensor(z_gm.dtype)
uint64_max = 2**64 - 1
mask = [uint64_max, uint64_max]
asc.add(z_local, x_local, y_local, mask=mask, repeat_times=1,
repeat_params=asc.BinaryRepeatParams(1, 1, 1, 8, 8, 8))
out_queue_z.enque(z_local)
in_queue_x.free_tensor(x_local)
in_queue_y.free_tensor(y_local)
@asc.jit
def copy_out(i: int, z_gm: asc.GlobalTensor, out_queue_z: asc.TQue, tile_length: asc.ConstExpr[int]):
z_local = out_queue_z.deque(z_gm.dtype)
asc.data_copy(z_gm[i * tile_length:], z_local, count=tile_length)
out_queue_z.free_tensor(z_local)
def vadd_launch(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
assert x.shape == y.shape
assert x.dtype == y.dtype
z = torch.zeros_like(x)
total_length = z.numel()
use_core_num = 16
block_length = (total_length + use_core_num - 1) // use_core_num
tile_length = 64
tile_num = (block_length + tile_length - 1) // tile_length
buffer_num = 1
vadd_kernel[use_core_num, rt.current_stream()](x, y, z, block_length, buffer_num, tile_length, tile_num)
return z
param_list = [
[torch.float32, (1000,)],
[torch.float32, (1,)],
[torch.float32, (9999,)],
[torch.float16, (2048,)],
[torch.int32, (8192,)],
[torch.int16, (8192,)],
[torch.int32, (153, 834)],
]
BACKENDS = [
config.Backend.NPU,
]
@pytest.mark.parametrize("dtype, size", param_list)
@pytest.mark.parametrize("backend", BACKENDS)
def test_vadd_bits(dtype, size, backend: config.Backend):
config.set_platform(backend)
device = "npu" if config.Backend(backend) == config.Backend.NPU else "cpu"
if dtype in {torch.float16, torch.float32}:
x = torch.randn(size, dtype=dtype, device=device)
y = torch.randn(size, dtype=dtype, device=device)
else:
x = torch.randint(-100, 99, size, dtype=dtype, device=device)
y = torch.randint(-100, 99, size, dtype=dtype, device=device)
z = vadd_launch(x, y)
assert torch.allclose(z, x + y)