【经验分享】寒武纪|【经验分享】寒武纪 pytorch-mlu 添加逐层算子方法
? 欢迎关注我的公众号 [极智视界],回复001获取Google编程规范
? O_o
? >_<
? o_O
? O_o
? ~_~
? o_O
? 本教程分享了在寒武纪设备上 pytorch-mlu 中添加逐层算子的方法。
? pytorch-mlu 逐层模式中算子间数据传递和存储的基本单元是 tensor。pytorch-mlu 根据 tensor 中的 device 属性值将算子分发到不同设备。以 abs() 算子为例,在 dispatch 阶段会根据 input_tensor 的设备属性值将算子调用分发到具体设备,逻辑如下图所示:
?Catch 通过注册添加 MLU 算子方式与 pytorch 源码解耦,下面介绍在 Catch 中添加 MLU 算子的具体步骤。
1、注册算子
?在 catch/torch_mlu/csrc/generated/aten_mlu_type_default.cpp
中注册算子:
.op(torch::RegisterOperators::options().schema("aten::add.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor")// NOLINT .impl_unboxedOnlyKernel(at::TensorTypeId::MLUTensorId)aliasAnalysis(c10::AliasAnalysisKind::FROM_SCHEMA))
2、算子分发 ? AtenMluType 和 AtenMluCustomType 是 Catch 模块中算子的入口。AtenMluType 类主要包含框架中的标准算子;而 AtenMluCustomType 类包含客制化的算子。根据算子属性选择在 AtenMluType 还是 AtenMluCustomType 中添加相应算子声明和实现。
- 标准算子分发
在catch/torch_mlu/csrc/aten/aten_mlu_type.h
和catch/torch_mlu/csrc/aten/aten_mlu_type.cpp
中添加算子声明和实现:
aten_mlu_type.h
static at::Tensor add(const at::Tensor& self, const at::Tensor& other, at::Scalar alpha);
aten_mlu_type.cpp
at::Tensor AtenMluType::add(const at::Tensor& self, const at::Tensor& other, at::Scalar alpha){
return OP_DISPATCH(add, self, other, alpha);
}
- 客制化算子分发
catch/torch_mlu/csrc/aten/aten_mlu_type.h
和 catch/torch_mlu/csrc/aten/aten_mlu_custom_type.cpp
中添加算子申明和实现:aten_mlu_type.h
static at::Tensor linear(const at::Tensor& input,
const at::Tensor& weight,
const at::Tensor& bias,
const at::Tensor& q_scale,
const at::Tensor& q_mode);
aten_mlu_custom_type.cpp
at::Tensor AtenMluCustomType::linear(const at::Tensor& input,
const at::Tensor& weight,
const at::Tensor& bias,
const at::Tensor& q_scale,
const at::Tensor& q_mode){
return OP_DISPATCH(linear, input, weight, bias, q_scale, q_mode);
}
3、修改 OpMethods 基类 ? 从 AtenMluType 和 AtenMluCustomType 中都会通过 OpMethods 下发到推理算子或训练算子。在
catch/torch_mlu/csrc/aten/operators/op_methods.h
和 catch/torch_mlu/csrc/aten/operators/op_methods.cpp
中添加算子申明和实现。OpMethods 中的实现部分为该算子的 CPU 实现。op_methods.h
virtual at::Tensor add(const at::Tensor& self, const at::Tensor& other, at::Scalar alpha);
op_methods.cpp
at::Tensor OpMethods::add(const at::Tensor& self,
const at::Tensor& other,
at::Scalar alpha){
auto input_cpu = self.cpu();
auto other_cpu = other.cpu();
auto output = at::add(input_cpu, other_cpu, alpha);
return output.to(at::Device(at::Device::Type::MLU));
}
4、下发算子 ? 在
catch/torch_mlu/csrc/aten/operators/cnml_ops.h
和 catch/torch_mlu/csrc/aten/operators/cnml_ops.cpp
中添加推理算子申明和实现。cnml_ops.h
at::Tensor add(const at::Tensor& self, const at::Tensor& other, at::Scalar alpha);
cnml_ops.cpp
at::Tensor CnmlOps::add(const at::Tensor& self, const at::Tensor& other, at::Scalar alpha){
CNML_DISPATCH(add, cnml_add, self, other, alpha);
// CNML_DISPATCH 宏第一个参数是该接口名,第二个参数是wrapper个名字,其余
}
5、添加 wrapper ? wrapper 是对算子 kernel 的封装,每个算子对应一个 wrapper。这里以 add 算子为例,添加 wrapper 如下所示:
cnml_kernel.h
at::Tensor cnml_add(const at::Tensor& input, const at::Tensor& other, at::Scalar alpha);
add.cpp
at::Tensor cnml_add(const at::Tensor& input, const at::Tensor& other, at::Scalar alpha_scalar){
TORCH_CHECK(input.dim() >= 0 || other.dim() >= 0, "dimension not support");
at::Tensor input_ = input;
at::Tensor other_ = other;
auto alpha_data = https://www.it610.com/article/alpha_scalar.to();
if(alpha_data != 1){
// scale_t
other_ = cnml::ops::cnml_scale(other_, alpha_data, 0);
}
if(other_.dim() < 1 && other_.device().type() == c10::DeviceType::CPU){
auto other_scalar = other_.item();
return cnml_add_internal(input_, other_scalar);
// 调用kernel
}
if(input_.dim() < 1 && input_.device().type() == c10::DeviceType::CPU){
auto input_scalar = input_.item();
return cnml_add_internal(other_, input_scalar);
// 调用 kernel
}bool broadcast = input_.sizes() != other_.sizes();
if(broadcast){
auto broadcast_size = at::infer_size(input.sizes(), other.sizes());
at::Tensor broadcast1 = cnml::ops::cnml_expand(input_, broadcast_size, false);
at::Tensor broadcast2 = cnml::ops::cnml_expand(other_, broadcast_size, false);
return cnml_add_internal(broadcast1, broadcast2);
// 调用 kernel
}else{
return cnml_add_internal(input_, other_);
//调用 kernel
}
return cnml_add_internal(input_, other_);
//调用 kernel
}
6、添加 wrapper ? Wrapper 中通过调用 kernel 实现算子功能。示例中调用的是 cnml_add_internal。算子的具体实现主要通过调用 CNML 库的接口来完成,下面是 CNML 库的逻辑:
?kernel 实现就是按照上述编程逻辑调用 CNML 库接口完成的,在
catch/torch_mlu/csrc/aten/operators/cnml/internal/cnml_internal.h
和 catch/torch_mlu/csrc/aten/operators/cnml/internal/add_internal/cpp
中添加 kernel 函数的声明和实现。cnml_internal.h
at::Tensor cnml_add_internal(const at::Tensor& input1, const at::Tensor& input2);
add_internal.cpp
at::Tensor cnml_add_internal(const at::Tensor& input1, const at::Tensor& input2){
auto output = at::native::empty_like(input1);
// prepare input cnml tensor
auto* input1_impl = getMluTensorImpl(input1);
// 获取MluTensorImpl
auto input1_cnml = input1_impl->CreateCnmlTensor(
CNML_TENSOR, toCnmlDataType(input1.dtype()));
// 类型自适应:toCnmlDataType()auto* input2_impl = getMluTensorImpl(input2);
auto input2_cnml = input2_impl->CreateCnmlTensor(
CNML_TENSOR, toCnmlDataType(input2.dtype()));
// prepare output cnml tensor
auto* output_impl = getMluTensorImpl(output);
auto output_cnml = output_impl->CreateCnmlTensor(
CNML_TENSOR, toCnmlDataType(output.dtype()));
// End the execution flow if not MLU device
CHECK_MLU_DEVICE(output);
// setup operator
cnmlBaseOp_t add_op;
TORCH_CNML_CHECK(cnmlCreateAddOp(&add_op, input1_cnml, input2_cnml, output_cnml));
// return to JIT if running mode is fuse
CHEXK_RETURN_TO_FUSE(add_op, output);
// compile op
TORCH_CNML_CHECK(cnmlCompileBaseOp(add_op, GET_CORE_VERSION, GET_CORE_NUMBER));
auto queue = getCurQueue();
TORCH_CNML_CHECK(cnmlComputeAddOpForward_V4(add_op,
NULL,
input1_impl->raw_mutable_data(),
NULL,
input2_impl->raw_mutable_data(),
NULL,
output_impl->raw_mutable_data(),
queue,
NULL));
syncQueue(queue);
TORCH_CNML_CHECK(cnmlDestroyBaseOp(&add_op));
return output;
}
- 对 MLU 不支持算子的处理
catch/torch_mlu/csrc/aten/operators/
目录下。op_methods.cpp
at::Tensor OpMethods::add(const at::Tensor& self,
const at::Tensor& other,
at::Scalar alpha){
auto input_cpu = self.cpu();
auto other_cpu = other.cpu();
auto output = at::add(input_cpu, other_cpu, alpha);
return output.to(at::Device(at::Device::Type::MLU));
}
- 对于新增的算子在执行过程中抛出异常时,如果 CPU 上没有对应的算子操作,那么该操作无法切换到 CPU 上运行;
- Wrapper一般以
cnml_
算子名命名,kernel一般以cnml_
算子名_internal
命名
def test_add(self):
# "Tensor + Tensor" mode testing
for shape1, shape2 in [((1,3,224,224),(1,3,224,224)),((2,30,80),(2,30,80)),((3,20),(3,20)),((10),(10))]:
input1_cpu = torch.rand(shape1, dtype=torch.float)
input2_cpu = torch.rand(shape2, dtype=torch.float)
input1_mlu = input1_cpu.to(xm.mlu_device())
input2_mlu = input2_cpu.to(xm.mlu_device())
# 在 CPU 上计算
output_cpu = input1_cpu + input2_cpu
# 在 MLU 上计算
output_mlu = input1_mlu + input2_mlu
# 计算 MLU 的误差,并确保相对误差在 2% 以内
self.assertTensorsEqual(output_cpu, output_mlu.cpu(), 0.02, use_MSE=True)
?以上分享了在寒武纪设备 pytorch-mlu 中添加逐层算子的方法,并以 add() 算子为例进行了示例编写,希望我的分享会对你的学习有一点帮助。
【【经验分享】寒武纪|【经验分享】寒武纪 pytorch-mlu 添加逐层算子方法】【公众号传送】
《【经验分享】寒武纪 pytorch-mlu 添加逐层算子方法》