西风逍遥游
4377e062dd
Squashed onto current main. Adds src/engine/common/Assert.h (EV_PARAM_CHECK / EV_ASSERT backed by zeroerr), an EVENGINE_ENABLE_ASSERTS option (non-Debug builds strip ZEROERR_NO_ASSERT unless enabled), parameter checks at key entry points, Doxygen comments across headers, Doxyfile + make docs / CMake docs targets, and Windows setup + code conventions in AGENTS.md.
204 lines
7.2 KiB
C++
204 lines
7.2 KiB
C++
#ifndef EVE_TENSOR_TENSOR_H
|
||
#define EVE_TENSOR_TENSOR_H
|
||
|
||
#include <cstdint>
|
||
#include <string>
|
||
#include <vector>
|
||
|
||
namespace eve::tensor {
|
||
|
||
class Graph;
|
||
class Func;
|
||
class TF;
|
||
|
||
/**
|
||
* @brief Tensor element types.
|
||
*
|
||
* Script-visible tensors are float32; int32 tensors are used for index data
|
||
* (argmax outputs, embedding lookups, cast("int32")). Int32 values are stored
|
||
* losslessly as floats for |v| < 2^24, which comfortably covers model
|
||
* vocabularies / sequence lengths / simulation ids used in games.
|
||
*/
|
||
enum class DType : uint8_t {
|
||
Float32 = 0,
|
||
Int32 = 1,
|
||
Fp16 = 2, // IEEE half, 2 bytes/elem, weight-only
|
||
Fp8E4M3 = 3, // 1 byte/elem, weight-only
|
||
Fp4E2M1 = 4, // 4-bit e2m1, two per byte, weight-only
|
||
Int8 = 5, // 1 byte/elem + per-group scale, weight-only
|
||
Int4 = 6, // 4-bit + per-group scale, weight-only
|
||
};
|
||
|
||
namespace q {
|
||
bool isQuantDType(DType dt);
|
||
}
|
||
|
||
const char *dtypeName(DType dtype);
|
||
bool parseDType(const std::string &name, DType &out);
|
||
|
||
/**
|
||
* @brief float32 / int32 tensor (rank 1–6), row-major.
|
||
* Eager: owns a buffer. Symbolic: node in a Func graph (no buffer until run).
|
||
*/
|
||
class Tensor {
|
||
public:
|
||
static constexpr int kMaxRank = 6;
|
||
|
||
Tensor() = default;
|
||
/** @brief 从 dims[0..rank) 创建 eager 张量。 */
|
||
explicit Tensor(const int *dims, int rank);
|
||
explicit Tensor(DType dtype, const int *dims, int rank);
|
||
/** @brief 按秩创建全零 eager 张量。 */
|
||
Tensor(int d0);
|
||
Tensor(int d0, int d1);
|
||
Tensor(int d0, int d1, int d2);
|
||
Tensor(int d0, int d1, int d2, int d3);
|
||
Tensor(int d0, int d1, int d2, int d3, int d4);
|
||
Tensor(int d0, int d1, int d2, int d3, int d4, int d5);
|
||
|
||
/** @brief Symbolic handle into a graph node. */
|
||
static Tensor *makeSymbolic(Graph *graph, int nodeId, const int *dims, int rank);
|
||
|
||
bool isSymbolic() const { return kind_ == Kind::Symbolic; }
|
||
bool isEager() const { return kind_ == Kind::Eager; }
|
||
Graph *graph() const { return graph_; }
|
||
int nodeId() const { return nodeId_; }
|
||
|
||
int getRank() const { return rank_; }
|
||
int getSize() const { return size_; }
|
||
int getDim(int axis) const;
|
||
int getDim0() const { return dims_[0]; }
|
||
int getDim1() const { return dims_[1]; }
|
||
int getDim2() const { return dims_[2]; }
|
||
int getDim3() const { return dims_[3]; }
|
||
int getDim4() const { return dims_[4]; }
|
||
int getDim5() const { return dims_[5]; }
|
||
std::string getDevice() const { return device_; }
|
||
std::string getDtype() const { return dtypeName(dtype_); }
|
||
DType dtype() const { return dtype_; }
|
||
void setDtype(DType dtype) { dtype_ = dtype; }
|
||
|
||
/** True for packed weight-quantization dtypes (bytes_, qScales_). */
|
||
bool isQuantized() const { return q::isQuantDType(dtype_); }
|
||
|
||
/** Dequantize this tensor to float32 (eager only). */
|
||
std::vector<float> dequantized() const;
|
||
|
||
/** Per-group scale vector + group size for quantized tensors. */
|
||
const std::vector<float> &qScales() const { return qScales_; }
|
||
int qGroup() const { return qGroup_; }
|
||
const std::vector<uint8_t> &qBytes() const { return bytes_; }
|
||
|
||
float get(int flatIndex) const;
|
||
void set(int flatIndex, float value);
|
||
float get1(int i0) const;
|
||
void set1(int i0, float value);
|
||
float get2(int i0, int i1) const;
|
||
void set2(int i0, int i1, float value);
|
||
float get3(int i0, int i1, int i2) const;
|
||
void set3(int i0, int i1, int i2, float value);
|
||
float get4(int i0, int i1, int i2, int i3) const;
|
||
void set4(int i0, int i1, int i2, int i3, float value);
|
||
float get5(int i0, int i1, int i2, int i3, int i4) const;
|
||
void set5(int i0, int i1, int i2, int i3, int i4, float value);
|
||
float get6(int i0, int i1, int i2, int i3, int i4, int i5) const;
|
||
void set6(int i0, int i1, int i2, int i3, int i4, int i5, float value);
|
||
|
||
void fill(float value);
|
||
void copyFrom(const Tensor *other);
|
||
Tensor *clone() const;
|
||
|
||
/** @brief Eager 逐元素运算（符号张量会抛异常）。 */
|
||
Tensor *add(const Tensor *other) const;
|
||
Tensor *sub(const Tensor *other) const;
|
||
Tensor *multiply(const Tensor *other) const;
|
||
Tensor *div(const Tensor *other) const;
|
||
Tensor *addScalar(float s) const;
|
||
Tensor *subScalar(float s) const;
|
||
Tensor *mulScalar(float s) const;
|
||
Tensor *divScalar(float s) const;
|
||
Tensor *neg() const;
|
||
Tensor *abs() const;
|
||
Tensor *sqrt() const;
|
||
Tensor *exp() const;
|
||
Tensor *log() const;
|
||
Tensor *sin() const;
|
||
Tensor *cos() const;
|
||
Tensor *tanh() const;
|
||
Tensor *relu() const;
|
||
Tensor *sigmoid() const;
|
||
Tensor *gelu() const;
|
||
Tensor *silu() const;
|
||
Tensor *powScalar(float exp) const;
|
||
Tensor *clamp(float lo, float hi) const;
|
||
Tensor *maximumScalar(float s) const;
|
||
Tensor *minimumScalar(float s) const;
|
||
|
||
/** @brief Eager 原地运算。 */
|
||
void addInPlace(const Tensor *other);
|
||
void multiplyInPlace(const Tensor *other);
|
||
void addScalarInPlace(float s);
|
||
void mulScalarInPlace(float s);
|
||
void reluInPlace();
|
||
|
||
/** @brief 归约：求和 / 均值 / 最小 / 最大。 */
|
||
float reduceSum() const;
|
||
float reduceMean() const;
|
||
float reduceMin() const;
|
||
float reduceMax() const;
|
||
float dot(const Tensor *other) const;
|
||
|
||
/** @brief 矩阵乘法 / 转置 / 变形。 */
|
||
Tensor *matmul(const Tensor *other) const;
|
||
Tensor *transpose() const;
|
||
Tensor *permute(const int *order, int rank) const;
|
||
Tensor *reshape1(int d0) const;
|
||
Tensor *reshape2(int d0, int d1) const;
|
||
Tensor *reshape3(int d0, int d1, int d2) const;
|
||
Tensor *reshape4(int d0, int d1, int d2, int d3) const;
|
||
Tensor *reshape5(int d0, int d1, int d2, int d3, int d4) const;
|
||
Tensor *reshape6(int d0, int d1, int d2, int d3, int d4, int d5) const;
|
||
Tensor *flatten() const;
|
||
|
||
/** @brief 原始数据指针（eager）。 */
|
||
float *data();
|
||
const float *data() const;
|
||
|
||
void ensureEager(const char *op) const;
|
||
|
||
static int product(const int *dims, int rank);
|
||
|
||
private:
|
||
friend class TF;
|
||
friend class Func;
|
||
friend class CompiledFunction;
|
||
friend class Graph;
|
||
|
||
enum class Kind { Eager, Symbolic };
|
||
|
||
void initDims(DType dtype, const int *dims, int rank);
|
||
void checkSameShape(const Tensor *other, const char *op) const;
|
||
int offset2(int i0, int i1) const;
|
||
int offset3(int i0, int i1, int i2) const;
|
||
int offset4(int i0, int i1, int i2, int i3) const;
|
||
int offset5(int i0, int i1, int i2, int i3, int i4) const;
|
||
int offset6(int i0, int i1, int i2, int i3, int i4, int i5) const;
|
||
|
||
Kind kind_ = Kind::Eager;
|
||
Graph *graph_ = nullptr;
|
||
int nodeId_ = -1;
|
||
int rank_ = 0;
|
||
int dims_[kMaxRank] = {0, 0, 0, 0, 0, 0};
|
||
int size_ = 0;
|
||
DType dtype_ = DType::Float32;
|
||
std::vector<float> data_;
|
||
std::vector<uint8_t> bytes_; // packed payload for quantized dtypes
|
||
std::vector<float> qScales_; // per-group scales (int8/int4)
|
||
int qGroup_ = 0; // elements per scale group
|
||
std::string device_ = "cpu";
|
||
};
|
||
|
||
} // namespace eve::tensor
|
||
|
||
#endif // EVE_TENSOR_TENSOR_H
|