vf.select

产品支持情况

• Ascend 950PR/Ascend 950DT：支持

• Atlas A3 训练系列产品/Atlas A3 推理系列产品：不支持

• Atlas A2 训练系列产品/Atlas A2 推理系列产品：不支持

功能说明

按掩码在两个输入间逐元素选择。掩码为真处取src0，为假处取src1。

\[\begin{split}dstReg_i = \begin{cases} srcTrueReg_i & \text{if } mask_i = 1 \\ srcFalseReg_i & \text{if } mask_i = 0 \end{cases}\end{split}\]

函数原型

select(src0, src1, preg) -> dst

参数说明

参数	输入/输出	说明
src0	输入	掩码为真时，取值的源操作数，为reg_tensor或者mask_reg类型，源操作数src0、src1与目的操作数dst的数据类型保持一致。支持的数据类型为：DT_BOOL、DT_INT8、DT_UINT8、DT_INT16、DT_UINT16、DT_FP16、DT_BF16、DT_INT32、DT_UINT32、DT_FP32、DT_INT64、DT_UINT64。
src1	输入	掩码为假时，取值的源操作数，为reg_tensor或者mask_reg类型，支持的数据类型与src0一致。
preg	输入	mask_reg。

约束说明

无。

返回值说明

返回dst目的操作数，为reg_tensor或者mask_reg类型，支持的数据类型与src0一致。

调用示例

reg_tensor调用示例

import os
import pypto_pro.language as pl
import torch
import torch_npu

@pl.vector_function
def example_vf(src_a, src_b, dst_tile):
    preg = vf.create_mask(pattern=pl.MaskPattern.ALL, dtype=pl.DT_FP32)
    reg_a = vf.load_align(src_a, 0)
    reg_b = vf.load_align(src_b, 0)
    dst_mask = vf.gt(reg_a, reg_b, preg)
    reg_out = vf.select(reg_a, reg_b, dst_mask)
    vf.store_align(dst_tile, reg_out, preg)

@pl.jit()
def example_kernel(
    a: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_FP32],
    b: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_FP32],
    out: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_FP32],
):
    tf = pl.TileType(shape=[1, 64], dtype=pl.DT_FP32, target_memory=pl.MemorySpace.Vec)
    in_a_grp = pl.make_tile_group(type=tf, addrs=0x0, mutex_ids=[0])
    in_a = in_a_grp.current()
    in_b_grp = pl.make_tile_group(type=tf, addrs=0x100, mutex_ids=[1])
    in_b = in_b_grp.current()
    t_out_grp = pl.make_tile_group(type=tf, addrs=0x200, mutex_ids=[2])
    t_out = t_out_grp.current()
    with pl.section_vector():
        pl.load(in_a, a, [0, 0])
        pl.load(in_b, b, [0, 0])
        example_vf(in_a, in_b, t_out)
        pl.store(out, t_out, [0, 0])

def test_example():
    device_id = int(os.environ.get("TILE_FWK_DEVICE_ID", 0))
    device = f"npu:{device_id}"
    core_nums = 1
    torch.npu.set_device(device)
    a = torch.randn([1, 64], device=device, dtype=torch.float32)
    b = torch.randn([1, 64], device=device, dtype=torch.float32)
    out = torch.empty([1, 64], device=device, dtype=torch.float32)
    example_kernel[None, core_nums](a, b, out)
    torch.npu.synchronize()
    torch.testing.assert_close(out, torch.where(a > b, a, b), rtol=1e-5, atol=1e-5)

if __name__ == "__main__":
    test_example()
    print("PASSED")

INT64数据类型示例

import os
import pypto_pro.language as pl
import torch
import torch_npu

@pl.vector_function
def example_vf_int64(src_a, src_b, dst_tile):
    preg = vf.create_mask(pattern=pl.MaskPattern.ALL, dtype=pl.DT_INT64)
    cmp_mask = vf.create_mask(pattern=pl.MaskPattern.ALL, dtype=pl.DT_INT64)
    reg_a = vf.load_align(src_a, 0)
    reg_b = vf.load_align(src_b, 0)
    gt_mask = vf.gt(reg_a, reg_b, cmp_mask)
    reg_out = vf.select(reg_a, reg_b, gt_mask)
    vf.store_align(dst_tile, reg_out, preg)

@pl.jit()
def example_kernel_int64(
    a: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_INT64],
    b: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_INT64],
    out: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_INT64],
):
    tf = pl.TileType(shape=[1, 32], dtype=pl.DT_INT64, target_memory=pl.MemorySpace.Vec)
    in_a_grp = pl.make_tile_group(type=tf, addrs=0, mutex_ids=[0])
    in_a = in_a_grp.current()
    in_b_grp = pl.make_tile_group(type=tf, addrs=256, mutex_ids=[1])
    in_b = in_b_grp.current()
    t_out_grp = pl.make_tile_group(type=tf, addrs=512, mutex_ids=[2])
    t_out = t_out_grp.current()
    with pl.section_vector():
        pl.load(in_a, a, [0, 0])
        pl.load(in_b, b, [0, 0])
        example_vf_int64(in_a, in_b, t_out)
        pl.store(out, t_out, [0, 0])

def test_example_int64():
    device_id = int(os.environ.get("TILE_FWK_DEVICE_ID", 0))
    device = f"npu:{device_id}"
    core_nums = 1
    torch.npu.set_device(device)
    a = torch.randint(-100, 100, [1, 32], device=device, dtype=torch.int64)
    b = torch.randint(-100, 100, [1, 32], device=device, dtype=torch.int64)
    out = torch.empty([1, 32], device=device, dtype=torch.int64)
    example_kernel_int64[None, core_nums](a, b, out)
    torch.npu.synchronize()
    torch.testing.assert_close(out, torch.where(a > b, a, b), rtol=0, atol=0)

if __name__ == "__main__":
    test_example_int64()
    print("PASSED")

mask_reg调用示例

当源操作数为mask_reg时，vf.select根据preg的比特位在两个mask_reg间选取。

import os
import pypto_pro.language as pl
import torch
import torch_npu

@pl.vector_function
def example_vf(src_tile, dst_tile):
    preg = vf.create_mask(pattern=pl.MaskPattern.ALL, dtype=pl.DT_FP32)
    reg = vf.load_align(src_tile, 0)
    mask_a = vf.ge(reg, 0.0, preg)
    mask_b = vf.lt(reg, 0.0, preg)
    preg_sel = vf.select(mask_a, mask_b, preg)
    reg_dst = vf.abs(reg, preg_sel)
    vf.store_align(dst_tile, reg_dst, preg)

@pl.jit()
def example_kernel(
    a: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_FP32],
    out: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_FP32],
):
    tf = pl.TileType(shape=[1, 64], dtype=pl.DT_FP32, target_memory=pl.MemorySpace.Vec)
    in_a_grp = pl.make_tile_group(type=tf, addrs=0x0, mutex_ids=[0])
    in_a = in_a_grp.current()
    t_out_grp = pl.make_tile_group(type=tf, addrs=0x100, mutex_ids=[1])
    t_out = t_out_grp.current()
    with pl.section_vector():
        pl.load(in_a, a, [0, 0])
        example_vf(in_a, t_out)
        pl.store(out, t_out, [0, 0])

def test_example():
    device_id = int(os.environ.get("TILE_FWK_DEVICE_ID", 0))
    device = f"npu:{device_id}"
    core_nums = 1
    torch.npu.set_device(device)
    a = torch.randn([1, 64], device=device, dtype=torch.float32)
    out = torch.empty([1, 64], device=device, dtype=torch.float32)
    example_kernel[None, core_nums](a, out)
    torch.npu.synchronize()
    expected = torch.where(a >= 0, a, torch.zeros_like(a))
    torch.testing.assert_close(out, expected, rtol=1e-5, atol=1e-5)

if __name__ == "__main__":
    test_example()
    print("PASSED")

INT64数据类型示例

import os
import pypto_pro.language as pl
import torch
import torch_npu

@pl.vector_function
def example_vf_int64(src_tile, dst_tile):
    preg = vf.create_mask(pattern=pl.MaskPattern.ALL, dtype=pl.DT_INT64)
    reg_a = vf.load_align(src_tile, 0)
    reg_out = vf.select(reg_a, reg_a, preg)
    vf.store_align(dst_tile, reg_out, preg)

@pl.jit()
def example_kernel_int64(
    a: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_INT64],
    out: pl.Tensor[[pl.DYNAMIC, pl.DYNAMIC], pl.DT_INT64],
):
    tf = pl.TileType(shape=[1, 32], dtype=pl.DT_INT64, target_memory=pl.MemorySpace.Vec)
    in_a_grp = pl.make_tile_group(type=tf, addrs=0, mutex_ids=[0])
    in_a = in_a_grp.current()
    t_out_grp = pl.make_tile_group(type=tf, addrs=256, mutex_ids=[1])
    t_out = t_out_grp.current()
    with pl.section_vector():
        pl.load(in_a, a, [0, 0])
        example_vf_int64(in_a, t_out)
        pl.store(out, t_out, [0, 0])

def test_example_int64():
    device_id = int(os.environ.get("TILE_FWK_DEVICE_ID", 0))
    device = f"npu:{device_id}"
    core_nums = 1
    torch.npu.set_device(device)
    a = torch.randint(-100, 100, [1, 32], device=device, dtype=torch.int64)
    out = torch.empty([1, 32], device=device, dtype=torch.int64)
    example_kernel_int64[None, core_nums](a, out)
    torch.npu.synchronize()
    torch.testing.assert_close(out, a, rtol=0, atol=0)

if __name__ == "__main__":
    test_example_int64()
    print("PASSED")