Skip to content

front&excuter:联合调试reshape #16

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
miaobyte merged 3 commits into from
Apr 6, 2025
Merged

front&excuter:联合调试reshape #16

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
fd47f3b
excuter(cpu/cuda):reshape修改为支持非inplace
miaobyte Apr 6, 2025
a501244
excuter(cpu/cuda):reshape修改为支持非inplace
miaobyte Apr 6, 2025
b9e7dc4
front&excuter:联合调试reshape
miaobyte Apr 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion doc/excuter/op-mem-cuda/list.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@
|-----------|--------|------------|--------------|----------------|
| concat | miaobyte | concat(listtensor<any> tensors, var<int32> axis)->(tensor<any> result) | Tresult = concat([T1, T2...], axis=3) | concat(listtensor<any> tensors, var<int32> axis)->(tensor<any> result) |
| transpose | miaobyte | transpose(tensor<any> A, vector<int32> dim_order)->(tensor<any> C) | T2 = T1.transpose(dimorder=[1,0]) | transpose(tensor<any> A, vector<int32> dim_order)->(tensor<any> C) |
| reshape | miaobyte | reshape(tensor<any> A, vector<int32> shape)->() | T2=T1.reshape(shape) | reshape(tensor<any> A, vector<int32> shape)->() |
| reshape | miaobyte | reshape(tensor<any> A, vector<int32> shape)->(tensor<any> B) | T1.reshape(shape)->T2 | reshape(tensor<any> A, vector<int32> shape)->(tensor<any> B) |
| matmul | cublas | matmul(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1 @ T2 | matmul(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| comparescalar | miaobyte | comparescalar(tensor<any> A, var<any> scalar)->(tensor<int8> mask) | mask=compare(T1, scalar) | comparescalar(tensor<any> A, var<any> scalar)->(tensor<int8> mask) |
| add | cublas | add(tensor<any> a, tensor<any> b)->(tensor<any> c) | T3=T1+T2 | add(tensor<any> a, tensor<any> b)->(tensor<any> c) |
Expand Down
6 changes: 3 additions & 3 deletions excuter/cpp-common/src/deepx/tensorfunc/changeshape.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,14 +11,14 @@ namespace deepx::tensorfunc
template <typename Author, typename T>
struct reshapeDispatcher
{
static void reshape(Tensor<T> &tensor, const std::vector<int> &new_shape) = delete;
static void reshape(const Tensor<T> &tensor, const std::vector<int> &new_shape,Tensor<T> &output) = delete;
};

// A.reshape(new_shape)
template <typename Author, typename T>
void reshape(Tensor<T> &tensor, const std::vector<int> &new_shape)
void reshape(const Tensor<T> &tensor, const std::vector<int> &new_shape,Tensor<T> &output)
{
reshapeDispatcher<Author, T>::reshape(tensor, new_shape);
reshapeDispatcher<Author, T>::reshape(tensor, new_shape,output);
}

template <typename Author, typename T>
Expand Down
2 changes: 1 addition & 1 deletion excuter/cpp-common/src/deepx/tf/tf.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,7 +110,7 @@ namespace deepx::tf
}
std::stringstream ss(textvalue);
std::string item;
while (std::getline(ss, item, ','))
while (std::getline(ss, item, ' '))
{
result.push_back(to<T>(item));
}
Expand Down
5 changes: 4 additions & 1 deletion excuter/op-mem-cuda/src/client/tfs.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -337,7 +337,10 @@ namespace deepx::tf
Param("A", DataCategory::Tensor, Precision::Any),
Param("shape", DataCategory::Vector, Precision::Int32),
}),
vector<Param>()));
vector<Param>(
{
Param("B", DataCategory::Tensor, Precision::Any),
})));

tffactory.add_tf(std::make_shared<Transpose<miaobyte>>(vector<Param>(
{
Expand Down
87 changes: 52 additions & 35 deletions excuter/op-mem-cuda/src/deepx/tensorfunc/changeshape_miaobyte.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,13 +14,30 @@ namespace deepx::tensorfunc
template <typename T>
struct reshapeDispatcher<miaobyte, T>
{
static void reshape(Tensor<T> &tensor, const std::vector<int> &new_shape)
static void reshape(const Tensor<T> &tensor, const std::vector<int> &shape, Tensor<T> &output)
{
if (tensor.shape.dim != new_shape.size())
int new_prod = 1;
for (int dim : shape)
{
throw std::runtime_error("Tensor shapes must match for reshape");
new_prod *= dim;
}

if (tensor.shape.size != new_prod)
{
throw std::invalid_argument("Shape size mismatch");
}
Shape newshape(shape);
if (tensor.data == output.data)
{
output.shape.shape=newshape.shape;
output.shape.strides=newshape.strides;
}
else
{
output.shape.shape=newshape.shape;
output.shape.strides=newshape.strides;
output.copyer(tensor.data, output.data, tensor.shape.size);
}
tensor.shape = Shape(new_shape);
}
};

Expand All @@ -35,46 +52,46 @@ namespace deepx::tensorfunc
}
auto [actual_blocks, optimal_block_size] = BestDims(tensor.shape.size);
launch_transpose<T>(actual_blocks, optimal_block_size,
tensor.data, tensor.shape.strides.data(),
output.data, output.shape.strides.data(),
tensor.shape.dim, tensor.shape.size, dim_order.data());
tensor.data, tensor.shape.strides.data(),
output.data, output.shape.strides.data(),
tensor.shape.dim, tensor.shape.size, dim_order.data());
}
};

template <typename T>
struct concatDispatcher<miaobyte, T>
{
static void concat(const vector<Tensor<T>*> tensors, const int axis, Tensor<T> &C)
{
//checkshape
if (!checkShapeConcat(tensors, axis, C))
{
throw TensorShapeError("Output tensor shape size must match the sum of input tensor shape sizes for concat");
}
static void concat(const vector<Tensor<T> *> tensors, const int axis, Tensor<T> &C)
{
// checkshape
if (!checkShapeConcat(tensors, axis, C))
{
throw TensorShapeError("Output tensor shape size must match the sum of input tensor shape sizes for concat");
}

vector<const T *> tensorsData(tensors.size());
for (int i = 0; i < tensors.size(); i++)
{
tensorsData[i] = tensors[i]->data;
}

vector<const T*> tensorsData(tensors.size());
for (int i = 0; i < tensors.size(); i++)
{
tensorsData[i] = tensors[i]->data;
}
vector<int> inputStrides;
for (int i = 0; i < tensors.size(); i++)
{
std::copy(tensors[i]->shape.strides.data(), tensors[i]->shape.strides.data() + tensors[i]->shape.dim, std::back_inserter(inputStrides));
}

vector< int> inputStrides;
for (int i = 0; i < tensors.size(); i++)
{
std::copy(tensors[i]->shape.strides.data(), tensors[i]->shape.strides.data() + tensors[i]->shape.dim, std::back_inserter(inputStrides));
}

vector<int> shapeAtAxis(tensors.size());
for (int i = 0; i < tensors.size(); i++)
{
shapeAtAxis[i] = tensors[i]->shape[axis];
}
vector<int> shapeAtAxis(tensors.size());
for (int i = 0; i < tensors.size(); i++)
{
shapeAtAxis[i] = tensors[i]->shape[axis];
}

launch_concat<T>(tensorsData.data(), inputStrides.data(),
C.data, C.shape.strides.data(),
C.shape.dim,
C.shape.size,
axis, tensors.size(), shapeAtAxis.data());
launch_concat<T>(tensorsData.data(), inputStrides.data(),
C.data, C.shape.strides.data(),
C.shape.dim,
C.shape.size,
axis, tensors.size(), shapeAtAxis.data());
};
};
}
Expand Down
14 changes: 7 additions & 7 deletions excuter/op-mem-cuda/src/deepx/tf/changeshape.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,7 @@ namespace deepx::tf

string math_formula() const override
{
return "T2=T1.reshape(shape)";
return "T1.reshape(shape)->T2";
}

shared_ptr<TF> clone() const override
Expand All @@ -47,22 +47,22 @@ namespace deepx::tf
switch (input_type)
{
case Precision::Float64:
reshape<Author, double>(*mem->gettensor<double>(this->args[0].textvalue), shape);
reshape<Author, double>(*mem->gettensor<double>(this->args[0].textvalue), shape, *mem->gettensor<double>(this->returns[0].textvalue));
break;
case Precision::Float32:
reshape<Author, float>(*mem->gettensor<float>(this->args[0].textvalue), shape);
reshape<Author, float>(*mem->gettensor<float>(this->args[0].textvalue), shape, *mem->gettensor<float>(this->returns[0].textvalue));
break;
case Precision::Int64:
reshape<Author, int64_t>(*mem->gettensor<int64_t>(this->args[0].textvalue), shape);
reshape<Author, int64_t>(*mem->gettensor<int64_t>(this->args[0].textvalue), shape, *mem->gettensor<int64_t>(this->returns[0].textvalue));
break;
case Precision::Int32:
reshape<Author, int32_t>(*mem->gettensor<int32_t>(this->args[0].textvalue), shape);
reshape<Author, int32_t>(*mem->gettensor<int32_t>(this->args[0].textvalue), shape, *mem->gettensor<int32_t>(this->returns[0].textvalue));
break;
case Precision::Int16:
reshape<Author, int16_t>(*mem->gettensor<int16_t>(this->args[0].textvalue), shape);
reshape<Author, int16_t>(*mem->gettensor<int16_t>(this->args[0].textvalue), shape, *mem->gettensor<int16_t>(this->returns[0].textvalue));
break;
case Precision::Int8:
reshape<Author, int8_t>(*mem->gettensor<int8_t>(this->args[0].textvalue), shape);
reshape<Author, int8_t>(*mem->gettensor<int8_t>(this->args[0].textvalue), shape, *mem->gettensor<int8_t>(this->returns[0].textvalue));
break;
default:
error = "Unsupported type: " + precision_str(input_type);
Expand Down
5 changes: 4 additions & 1 deletion excuter/op-mem-ompsimd/src/client/tfs.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -328,7 +328,10 @@ namespace deepx::tf
Param("A", DataCategory::Tensor, Precision::Any),
Param("shape", DataCategory::Vector, Precision::Int32),
}),
vector<Param>()));
vector<Param>(
{
Param("B", DataCategory::Tensor, Precision::Any),
})));

tffactory.add_tf(std::make_shared<Transpose<miaobyte>>(vector<Param>(
{
Expand Down
17 changes: 14 additions & 3 deletions excuter/op-mem-ompsimd/src/deepx/tensorfunc/changeshape_miaobyte.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,9 +15,9 @@ namespace deepx::tensorfunc
template <typename T>
struct reshapeDispatcher<miaobyte, T>
{
static void reshape(Tensor<T> &tensor, const std::vector<int> &shape)
static void reshape(const Tensor<T> &tensor, const std::vector<int> &shape,Tensor<T> &output)
{ // 参数改为单个tensor引用

int new_prod = 1;
for (int dim : shape)
{
Expand All @@ -28,7 +28,18 @@ namespace deepx::tensorfunc
{
throw std::invalid_argument("Shape size mismatch");
}
tensor.shape = Shape(shape);
Shape newshape(shape);
if (tensor.data == output.data)
{
output.shape.shape=newshape.shape;
output.shape.strides=newshape.strides;
}
else
{
output.shape.shape=newshape.shape;
output.shape.strides=newshape.strides;
output.copyer(tensor.data, output.data, tensor.shape.size);
}
}
};

Expand Down
16 changes: 8 additions & 8 deletions excuter/op-mem-ompsimd/src/deepx/tf/changeshape.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ namespace deepx::tf

string math_formula() const override
{
return "T2=T1.reshape(shape)";
return "T1.reshape(shape)->T2";
}

shared_ptr<TF> clone() const override
Expand All @@ -46,22 +46,22 @@ namespace deepx::tf
switch (input_type)
{
case Precision::Float64:
reshape<Author, double>(*mem->gettensor<double>(this->args[0].textvalue), shape);
reshape<Author, double>(*mem->gettensor<double>(this->args[0].textvalue), shape, *mem->gettensor<double>(this->returns[0].textvalue));
break;
case Precision::Float32:
reshape<Author, float>(*mem->gettensor<float>(this->args[0].textvalue), shape);
reshape<Author, float>(*mem->gettensor<float>(this->args[0].textvalue), shape, *mem->gettensor<float>(this->returns[0].textvalue));
break;
case Precision::Int64:
reshape<Author, int64_t>(*mem->gettensor<int64_t>(this->args[0].textvalue), shape);
reshape<Author, int64_t>(*mem->gettensor<int64_t>(this->args[0].textvalue), shape, *mem->gettensor<int64_t>(this->returns[0].textvalue));
break;
case Precision::Int32:
reshape<Author, int32_t>(*mem->gettensor<int32_t>(this->args[0].textvalue), shape);
reshape<Author, int32_t>(*mem->gettensor<int32_t>(this->args[0].textvalue), shape, *mem->gettensor<int32_t>(this->returns[0].textvalue));
break;
case Precision::Int16:
reshape<Author, int16_t>(*mem->gettensor<int16_t>(this->args[0].textvalue), shape);
reshape<Author, int16_t>(*mem->gettensor<int16_t>(this->args[0].textvalue), shape, *mem->gettensor<int16_t>(this->returns[0].textvalue));
break;
case Precision::Int8:
reshape<Author, int8_t>(*mem->gettensor<int8_t>(this->args[0].textvalue), shape);
reshape<Author, int8_t>(*mem->gettensor<int8_t>(this->args[0].textvalue), shape, *mem->gettensor<int8_t>(this->returns[0].textvalue));
break;
default:
error = "Unsupported type: " + precision_str(input_type);
Expand All @@ -85,7 +85,7 @@ namespace deepx::tf

string math_formula() const override
{
return "T2 = T1.transpose(dimorder=[1,0])";
return "T1.transpose(dimorder=[1,0])->T2";
}

shared_ptr<TF> clone() const override
Expand Down
4 changes: 2 additions & 2 deletions front/py/deepx/nn/deepxir.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,8 +40,8 @@ def __init__(self,
"""

self._name = name
self._args = args
self._returns = returns
self._args = [arg if isinstance(arg, Param) else Param(arg) for arg in args]
self._returns = [ret if isinstance(ret, Param) else Param(ret) for ret in returns]
self._author = author
self._id=None
self._created_at=time.time()
Expand Down
4 changes: 2 additions & 2 deletions front/py/deepx/nn/functional/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
from .matmul import matmul
from .init import *
from .reduce import reduce_max,reduce_min,sum,prod,mean
from .changeshape import transpose,reshape,broadcast_shape,broadcast_to,unsqueeze
from .changeshape import *
from .activite import relu,sigmoid,swish
__all__ = [
"newtensor",
Expand All @@ -13,7 +13,7 @@
"add","sub","mul","div","sqrt","pow","exp","log","rsqrt",
"matmul",
"max","min","sum","prod","mean",
"transpose","reshape","broadcast_shape","broadcast_to","unsqueeze",
"reshape","permute","transpose",
"relu","sigmoid","swish",

]
Loading