Skip to content

excuter(cpu/cuda):reshape,transpose,concat #15

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 8 commits into from
Apr 6, 2025
Merged

excuter(cpu/cuda):reshape,transpose,concat #15

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
ea84ac1
fp16&bf16:cuda版本修正。12.1->12.6
miaobyte Apr 3, 2025
0830944
tf:todo清单
miaobyte Apr 3, 2025
b2fce25
excuter(cpu/cuda):reshape,transpose,concat
miaobyte Apr 5, 2025
737239b
excuter(cpu/cuda):reshape,transpose
miaobyte Apr 6, 2025
731b833
excuter(cpu/cuda):reshape,transpose,concat
miaobyte Apr 6, 2025
72e5334
front:Apply fix,requires_grad=requires_grad
miaobyte Apr 6, 2025
fd2ed6a
front:Apply
miaobyte Apr 6, 2025
eed8fce
Merge branch 'main' into main
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
4 changes: 4 additions & 0 deletions doc/excuter/op-mem-cuda/list.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,11 @@

| Operation | Author | Func Def | Math Formula | IR Instruction |
|-----------|--------|------------|--------------|----------------|
| 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)->() |
| 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) |
| add | miaobyte | add(tensor<any> a, tensor<any> b)->(tensor<any> c) | T3=T1+T2 | add(tensor<any> a, tensor<any> b)->(tensor<any> c) |
| uniform | miaobyte | uniform(tensor<any> t, var<any> low, var<any> high, var<int32> seed)->() | uniform(T1,low,high,seed) | uniform(tensor<any> t, var<any> low, var<any> high, var<int32> seed)->() |
Expand Down
59 changes: 31 additions & 28 deletions doc/excuter/op-mem-ompsimd/list.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,35 +4,38 @@

| Operation | Author | Func Def | Math Formula | IR Instruction |
|-----------|--------|------------|--------------|----------------|
| concat | none | concat()->() | Tresult = concat([T1, T2...], axis=3) | concat()->() |
| matmul | cblas | matmul(tensor<float64|float32> A, tensor<float64|float32> B)->(tensor<float64|float32> C) | T3=T1 @ T2 | matmul(tensor<float64|float32> A, tensor<float64|float32> B)->(tensor<float64|float32> C) |
| matmul | miaobyte | matmul(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1 @ T2 | matmul(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| compare | miaobyte | compare(tensor<any> A, tensor<any> B)->(tensor<int8> mask) | mask=compare(T1,T2) | compare(tensor<any> A, tensor<any> B)->(tensor<int8> mask) |
| min | miaobyte | min(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=min(T1,T2) | min(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| minscalar | miaobyte | minscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=min(T1,scalar) | minscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| exp | miaobyte | exp(tensor<any> A)->(tensor<any> C) | T3=exp(T1) | exp(tensor<any> A)->(tensor<any> C) |
| maxscalar | miaobyte | maxscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=max(T1,scalar) | maxscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| pow | miaobyte | pow(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1^T2 | pow(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| powscalar | miaobyte | powscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=T1^scalar | powscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| rdivscalar | miaobyte | rdivscalar(var<any> scalar, tensor<any> A)->(tensor<any> C) | T3=scalar/T1 | rdivscalar(var<any> scalar, tensor<any> A)->(tensor<any> C) |
| div | miaobyte | div(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1/T2 | div(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| sub | miaobyte | sub(tensor<any> a, tensor<any> b)->(tensor<any> c) | T3=T1-T2 | sub(tensor<any> a, tensor<any> b)->(tensor<any> c) |
| argset | none | argset(var<any> value)->(var<any> name) | var argname = argvalue | argset(var<any> value)->(var<any> name) |
| mulscalar | miaobyte | mulscalar(tensor<any> A, var<any> b)->(tensor<any> C) | T3=T1*scalar | mulscalar(tensor<any> A, var<any> b)->(tensor<any> C) |
| sqrt | miaobyte | sqrt(tensor<any> A)->(tensor<any> C) | T3=sqrt(T1) | sqrt(tensor<any> A)->(tensor<any> C) |
| vecset | none | vecset(vector<any> value)->(vector<any> name) | shape = [3 4 5] | vecset(vector<any> value)->(vector<any> name) |
| newtensor | none | newtensor(vector<int32> shape)->(tensor<any> tensor1) | T1 =Tensor(shape=[...]) | newtensor(vector<int32> shape)->(tensor<any> tensor1) |
| newtensor | none | newtensor(var<string> shape)->(tensor<any> tensor1) | T1 =Tensor(shape=[...]) | newtensor(var<string> shape)->(tensor<any> tensor1) |
| print | miaobyte | print(tensor<any> )->() | print(T1) | print(tensor<any> )->() |
| print | miaobyte | print(tensor<any> , var<string> )->() | print(T1) | print(tensor<any> , var<string> )->() |
| max | miaobyte | max(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=max(T1,T2) | max(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| divscalar | miaobyte | divscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=T1/scalar | divscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| constant | miaobyte | constant(tensor<any> t, var<any> value)->() | constant(T1,value) | constant(tensor<any> t, var<any> value)->() |
| arange | miaobyte | arange(tensor<any> t, var<any> start, var<any> step)->() | arange(T1,start,step) | arange(tensor<any> t, var<any> start, var<any> step)->() |
| subscalar | miaobyte | subscalar(tensor<any> a, var<any> scalar)->(tensor<any> c) | T3=T1-scalar | subscalar(tensor<any> a, var<any> scalar)->(tensor<any> c) |
| log | miaobyte | log(tensor<any> A)->(tensor<any> C) | T3=log(T1) | log(tensor<any> A)->(tensor<any> C) |
| uniform | miaobyte | uniform(tensor<any> t, var<any> low, var<any> high, var<int32> seed)->() | uniform(T1,low,high,seed) | uniform(tensor<any> t, var<any> low, var<any> high, var<int32> seed)->() |
| 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) |
| add | cblas | add(tensor<float64|float32> a, tensor<float64|float32> b)->(tensor<float64|float32> c) | T3=T1+T2 | add(tensor<float64|float32> a, tensor<float64|float32> b)->(tensor<float64|float32> c) |
| add | miaobyte | add(tensor<any> a, tensor<any> b)->(tensor<any> c) | T3=T1+T2 | add(tensor<any> a, tensor<any> b)->(tensor<any> c) |
| comparescalar | miaobyte | comparescalar(tensor<any> A, var<any> scalar)->(tensor<float32> mask) | mask=compare(T1,scalar) | comparescalar(tensor<any> A, var<any> scalar)->(tensor<float32> mask) |
| uniform | miaobyte | uniform(tensor<any> t, var<any> low, var<any> high, var<int32> seed)->() | uniform(T1,low,high,seed) | uniform(tensor<any> t, var<any> low, var<any> high, var<int32> seed)->() |
| addscalar | miaobyte | addscalar(tensor<any> a, var<any> scalar)->(tensor<any> c) | T3=T1+scalar | addscalar(tensor<any> a, var<any> scalar)->(tensor<any> c) |
| log | miaobyte | log(tensor<any> A)->(tensor<any> C) | T3=log(T1) | log(tensor<any> A)->(tensor<any> C) |
| reshape | miaobyte | reshape(tensor<any> A, vector<int32> shape)->() | T2=T1.reshape(shape) | reshape(tensor<any> A, vector<int32> shape)->() |
| arange | miaobyte | arange(tensor<any> t, var<any> start, var<any> step)->() | arange(T1,start,step) | arange(tensor<any> t, var<any> start, var<any> step)->() |
| divscalar | miaobyte | divscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=T1/scalar | divscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| print | miaobyte | print(tensor<any> )->() | print(T1) | print(tensor<any> )->() |
| print | miaobyte | print(tensor<any> , var<string> )->() | print(T1) | print(tensor<any> , var<string> )->() |
| newtensor | none | newtensor(vector<int32> shape)->(tensor<any> tensor1) | T1 =Tensor(shape=[...]) | newtensor(vector<int32> shape)->(tensor<any> tensor1) |
| newtensor | none | newtensor(var<string> shape)->(tensor<any> tensor1) | T1 =Tensor(shape=[...]) | newtensor(var<string> shape)->(tensor<any> tensor1) |
| vecset | none | vecset(vector<any> value)->(vector<any> name) | shape = [3 4 5] | vecset(vector<any> value)->(vector<any> name) |
| subscalar | miaobyte | subscalar(tensor<any> a, var<any> scalar)->(tensor<any> c) | T3=T1-scalar | subscalar(tensor<any> a, var<any> scalar)->(tensor<any> c) |
| sqrt | miaobyte | sqrt(tensor<any> A)->(tensor<any> C) | T3=sqrt(T1) | sqrt(tensor<any> A)->(tensor<any> C) |
| argset | none | argset(var<any> value)->(var<any> name) | var argname = argvalue | argset(var<any> value)->(var<any> name) |
| sub | miaobyte | sub(tensor<any> a, tensor<any> b)->(tensor<any> c) | T3=T1-T2 | sub(tensor<any> a, tensor<any> b)->(tensor<any> c) |
| mulscalar | miaobyte | mulscalar(tensor<any> A, var<any> b)->(tensor<any> C) | T3=T1*scalar | mulscalar(tensor<any> A, var<any> b)->(tensor<any> C) |
| div | miaobyte | div(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1/T2 | div(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| constant | miaobyte | constant(tensor<any> t, var<any> value)->() | constant(T1,value) | constant(tensor<any> t, var<any> value)->() |
| powscalar | miaobyte | powscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=T1^scalar | powscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| max | miaobyte | max(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=max(T1,T2) | max(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| pow | miaobyte | pow(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1^T2 | pow(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| maxscalar | miaobyte | maxscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=max(T1,scalar) | maxscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| mul | miaobyte | mul(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1*T2 | mul(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| exp | miaobyte | exp(tensor<any> A)->(tensor<any> C) | T3=exp(T1) | exp(tensor<any> A)->(tensor<any> C) |
| rdivscalar | miaobyte | rdivscalar(var<any> scalar, tensor<any> A)->(tensor<any> C) | T3=scalar/T1 | rdivscalar(var<any> scalar, tensor<any> A)->(tensor<any> C) |
| minscalar | miaobyte | minscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) | T3=min(T1,scalar) | minscalar(tensor<any> A, var<any> scalar)->(tensor<any> C) |
| min | miaobyte | min(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=min(T1,T2) | min(tensor<any> A, tensor<any> B)->(tensor<any> C) |
| compare | miaobyte | compare(tensor<any> A, tensor<any> B)->(tensor<float32> mask) | mask=compare(T1,T2) | compare(tensor<any> A, tensor<any> B)->(tensor<float32> mask) |
| matmul | cblas | matmul(tensor<float64|float32> A, tensor<float64|float32> B)->(tensor<float64|float32> C) | T3=T1 @ T2 | matmul(tensor<float64|float32> A, tensor<float64|float32> B)->(tensor<float64|float32> C) |
| matmul | miaobyte | matmul(tensor<any> A, tensor<any> B)->(tensor<any> C) | T3=T1 @ T2 | matmul(tensor<any> A, tensor<any> B)->(tensor<any> C) |
25 changes: 22 additions & 3 deletions excuter/cpp-common/src/deepx/dtype.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,9 +3,30 @@

#include <string>
#include <cstdint>
#include <sstream>

namespace deepx
{
template <typename T>
T to(const std::string &textvalue)
{
if constexpr (std::is_same_v<T, std::string>)
{
return textvalue;
}
else if constexpr (std::is_arithmetic_v<T>)
{
return static_cast<T>(std::stof(textvalue));
}
else
{
// 对于其他类型，尝试从字符串转换
T value;
std::istringstream iss(textvalue);
iss >> value;
return value;
}
}

enum class DataCategory : uint8_t
{
Expand Down Expand Up @@ -112,7 +133,7 @@ namespace deepx
// 布尔类型 (13位)
Bool = 1 << 13, // 0010 0000 0000 0000
String = 1 << 15, // 0100 0000 0000 0000
// 常用组合
// 常用组合
Any = 0xFFFF, // 1111 1111 1111 1111
Float = Float64 | Float32 | Float16 | BFloat16 | Float8E5M2 | Float8E4M3 | Float4E2M1,
Float8 = Float8E5M2 | Float8E4M3, // 所有FP8格式
Expand Down Expand Up @@ -230,8 +251,6 @@ namespace deepx
return TypeDef(category, precision);
}



// 修改precision_str函数以使用标准命名格式
inline std::string precision_str(Precision p)
{
Expand Down
1 change: 1 addition & 0 deletions excuter/cpp-common/src/deepx/mem/mem.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,6 +150,7 @@ namespace deepx::mem

return tensors;
}


void delete_tensor(const string &name)
{
Expand Down
31 changes: 29 additions & 2 deletions excuter/cpp-common/src/deepx/shape_concat.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@

#include "deepx/shape.hpp"
#include "deepx/tensor.hpp"
#include "stdutil/error.hpp"

namespace deepx
{
Expand All @@ -18,6 +19,32 @@ namespace deepx
}
return concatShape(shapes,axis);
}
}

#endif
template<typename T>
bool checkShapeConcat(const std::vector<Tensor<T>*> &tensors,const int axis,const Tensor<T> &output){
int axisDim=0;
for (int i = 0; i < tensors.size(); i++)
{
if (tensors[i]->shape.dim != output.shape.dim)
{
throw TensorShapeError("All input tensors must have the same dimension size for concat");
}
for (int j = 0; j < tensors[i]->shape.dim; j++)
{
if (j != axis)
{
if (tensors[i]->shape[j] != output.shape[j])
{
throw TensorShapeError("All input tensors must have the same dimension size for concat");
}
}
else
{
axisDim += tensors[i]->shape[j];
}
}
}
return axisDim == output.shape[axis];
}
};
#endif // DEEPX_SHAPE_CONCAT_HPP
150 changes: 117 additions & 33 deletions excuter/cpp-common/src/deepx/tensorfunc/changeshape.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,48 +1,132 @@
#ifndef DEEPX_TENSORFUNC_CHANGE_SHAPE_HPP
#define DEEPX_TENSORFUNC_CHANGE_SHAPE_HPP
#ifndef DEEPX_TENSORFUNC_CHANGESHAPE_HPP
#define DEEPX_TENSORFUNC_CHANGESHAPE_HPP

#include <vector>
#include "deepx/tensor.hpp"
#include "stdutil/error.hpp"

namespace deepx::tensorfunc
{
using namespace std;
template <typename Author, typename T>
struct reshapeDispatcher
{
static void reshape(Tensor<T> &tensor, const std::vector<int> &new_shape) = delete;
};

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

template <typename Author, typename T>
struct transposeDispatcher
{
static void transpose(const Tensor<T> &tensor, const std::vector<int> &dim_order, Tensor<T> &output) = delete;
};

// transpose(A,dim_order)=>B
template <typename Author, typename T>
void transpose(const Tensor<T> &tensor, const std::vector<int> &dim_order, Tensor<T> &output)
{
transposeDispatcher<Author, T>::transpose(tensor, dim_order, output);
}

template <typename Author, typename T>
struct concatDispatcher
{
static void concat(const vector<Tensor<T>*> tensors, const int axis, Tensor<T> &C) = delete;
};
// concat(tensors,axis)=>C
template <typename Author, typename T>
void concat(const vector<Tensor<T>*> tensors, const int axis, Tensor<T> &C)
{
concatDispatcher<Author, T>::concat(tensors, axis, C);
}

// https://onnx.ai/onnx/operators/onnx__Split.html
template <typename Author, typename T>
struct splitDispatcher
{
static void split(const Tensor<T> &A, const int axis,const std::vector<int> &splits, Tensor<T> *&B) = delete;
static void split(const Tensor<T> &A, const int axis,const int num_outputs, Tensor<T> *&B) = delete;
};
// split(tensor,axis,splits)=>tensors
template <typename Author, typename T>
void split(const Tensor<T> &A, const int axis,const std::vector<int> &splits, Tensor<T> *&B)
{
splitDispatcher<Author, T>::split(A, axis, splits, B);

}

// split(tensor,axis,num_outputs)=>tensors
template <typename Author, typename T>
void split(const Tensor<T> &A, const int axis,const int num_outputs, Tensor<T> *&B)
{
splitDispatcher<Author, T>::split(A, axis, num_outputs, B);
}

// 通用模板声明
template <typename Author, typename T>
struct InitDispatcher
struct expandDispatcher
{
static void reshape(Tensor<T> &tensor, const Shape &new_shape) = delete;
static void expand(const Tensor<T> &A, const Shape &new_shape, Tensor<T> &B) = delete;
};

template <typename Author, typename T>
void reshape(Tensor<T> &tensor, const Shape &new_shape)
{
InitDispatcher<Author, T>::reshape(tensor, new_shape);
}

// // 作者特化示例（类型无关实现）
// template <typename T>
// struct InitDispatcher<miaobyte, T>
// {
// static void reshape(Tensor<T> &tensor, const Shape &new_shape)
// {
// // 统一实现，不依赖T的类型
// if (tensor.shape.size() != new_shape.size())
// {
// throw std::invalid_argument("Total elements must match");
// }
// tensor.shape = new_shape;
// }
// };
// 特化作者和具体精度
// template <>
// struct InitDispatcher<miaobyte, float>
// {
// static void reshape(Tensor<float> &tensor, const Shape &new_shape)
// {
// // CUDA实现
// }
// };
void expand(const Tensor<T> &A, const Shape &new_shape, Tensor<T> &B)
{
expandDispatcher<Author, T>::expand(A, new_shape, B);
}

template <typename Author, typename T>
struct squeezeDispatcher
{
static void squeeze(Tensor<T> &tensor) = delete;
};

template <typename Author, typename T>
void squeeze(Tensor<T> &tensor)
{
squeezeDispatcher<Author, T>::squeeze(tensor);
}

template <typename Author, typename T>
struct unsqueezeDispatcher
{
static void unsqueeze(Tensor<T> &tensor, const int axis) = delete;
};

template <typename Author, typename T>
void unsqueeze(Tensor<T> &tensor, const int axis)
{
unsqueezeDispatcher<Author, T>::unsqueeze(tensor, axis);
}

template <typename Author, typename T>
struct flattenDispatcher
{
static void flatten(Tensor<T> &tensor) = delete;
};

template <typename Author, typename T>
void flatten(Tensor<T> &tensor)
{
flattenDispatcher<Author, T>::flatten(tensor);
}

template <typename Author, typename T>
struct paddingDispatcher
{
static void padding(Tensor<T> &tensor, const Shape &new_shape) = delete;
};

template <typename Author, typename T>
void padding(Tensor<T> &tensor, const Shape &new_shape)
{
paddingDispatcher<Author, T>::padding(tensor, new_shape);
}
}

#endif
Loading