Commit 927b7781d935f4953351e0436717bc3d124040a3
1 parent
da2c7725
exposed jitcv Matrix class
Showing
5 changed files
with
268 additions
and
147 deletions
sdk/jitcv/jitcv.cpp
0 → 100644
| 1 | +#include "jitcv.h" | ... | ... |
sdk/jitcv/jitcv.h
0 → 100644
| 1 | +#ifndef __JITCV_H | |
| 2 | +#define __JITCV_H | |
| 3 | + | |
| 4 | +#include <stddef.h> | |
| 5 | +#include <stdint.h> | |
| 6 | + | |
| 7 | +namespace jitcv | |
| 8 | +{ | |
| 9 | + | |
| 10 | +/*! | |
| 11 | + * \brief jitcv matrix | |
| 12 | + * \author Josh Klontz \cite jklontz | |
| 13 | + * \note Not part of the core SDK | |
| 14 | + */ | |
| 15 | +struct Matrix | |
| 16 | +{ | |
| 17 | + uint8_t *data; /*!< Data */ | |
| 18 | + uint32_t channels; /*!< Channels */ | |
| 19 | + uint32_t columns; /*!< Columns */ | |
| 20 | + uint32_t rows; /*!< Rows */ | |
| 21 | + uint32_t frames; /*!< Frames */ | |
| 22 | + uint16_t hash; /*!< Bits : 8 | |
| 23 | + Floating : 1 | |
| 24 | + Signed : 1 | |
| 25 | + (Reserved) : 2 | |
| 26 | + Single-channel : 1 | |
| 27 | + Single-column : 1 | |
| 28 | + Single-row : 1 | |
| 29 | + Single-frame : 1 */ | |
| 30 | + | |
| 31 | + enum Hash { Bits = 0x00FF, | |
| 32 | + Floating = 0x0100, | |
| 33 | + Signed = 0x0200, | |
| 34 | + SingleChannel = 0x1000, | |
| 35 | + SingleColumn = 0x2000, | |
| 36 | + SingleRow = 0x4000, | |
| 37 | + SingleFrame = 0x8000, | |
| 38 | + u1 = 1, | |
| 39 | + u8 = 8, | |
| 40 | + u16 = 16, | |
| 41 | + u32 = 32, | |
| 42 | + u64 = 64, | |
| 43 | + s8 = 8 + Signed, | |
| 44 | + s16 = 16 + Signed, | |
| 45 | + s32 = 32 + Signed, | |
| 46 | + s64 = 64 + Signed, | |
| 47 | + f16 = 16 + Floating + Signed, | |
| 48 | + f32 = 32 + Floating + Signed, | |
| 49 | + f64 = 64 + Floating + Signed }; | |
| 50 | + | |
| 51 | + Matrix() : data(NULL), channels(0), columns(0), rows(0), frames(0), hash(0) {} | |
| 52 | + | |
| 53 | + Matrix(uint32_t _channels, uint32_t _columns, uint32_t _rows, uint32_t _frames, uint16_t _hash) | |
| 54 | + : data(NULL), channels(_channels), columns(_columns), rows(_rows), frames(_frames), hash(_hash) | |
| 55 | + { | |
| 56 | + setSingleChannel(channels == 1); | |
| 57 | + setSingleColumn(columns == 1); | |
| 58 | + setSingleRow(rows == 1); | |
| 59 | + setSingleFrame(frames == 1); | |
| 60 | + } | |
| 61 | + | |
| 62 | + inline void copyHeader(const Matrix &other) { channels = other.channels; columns = other.columns; rows = other.rows; frames = other.frames; hash = other.hash; } | |
| 63 | + inline void allocate() { deallocate(); data = new uint8_t[bytes()]; } | |
| 64 | + inline void deallocate() { delete[] data; data = NULL; } | |
| 65 | + | |
| 66 | + inline int bits() const { return hash & Bits; } | |
| 67 | + inline void setBits(int bits) { hash &= ~Bits; hash |= bits & Bits; } | |
| 68 | + inline bool isFloating() const { return hash & Floating; } | |
| 69 | + inline void setFloating(bool isFloating) { isFloating ? setSigned(true), hash |= Floating : hash &= ~Floating; } | |
| 70 | + inline bool isSigned() const { return hash & Signed; } | |
| 71 | + inline void setSigned(bool isSigned) { isSigned ? hash |= Signed : hash &= ~Signed; } | |
| 72 | + inline int type() const { return hash & (Bits + Floating + Signed); } | |
| 73 | + inline void setType(int type) { hash &= ~(Bits + Floating + Signed); hash |= type & (Bits + Floating + Signed); } | |
| 74 | + inline bool singleChannel() const { return hash & SingleChannel; } | |
| 75 | + inline void setSingleChannel(bool singleChannel) { singleChannel ? hash |= SingleChannel : hash &= ~SingleChannel; } | |
| 76 | + inline bool singleColumn() const { return hash & SingleColumn; } | |
| 77 | + inline void setSingleColumn(bool singleColumn) { singleColumn ? hash |= SingleColumn : hash &= ~SingleColumn; } | |
| 78 | + inline bool singleRow() const { return hash & SingleRow; } | |
| 79 | + inline void setSingleRow(bool singleRow) { singleRow ? hash |= SingleRow : hash &= ~SingleRow; } | |
| 80 | + inline bool singleFrame() const { return hash & SingleFrame; } | |
| 81 | + inline void setSingleFrame(bool singleFrame) { singleFrame ? hash |= SingleFrame : hash &= ~SingleFrame; } | |
| 82 | + inline uint32_t elements() const { return channels * columns * rows * frames; } | |
| 83 | + inline uint32_t bytes() const { return bits() / 8 * elements(); } | |
| 84 | +}; | |
| 85 | + | |
| 86 | +} // namespace jitcv | |
| 87 | + | |
| 88 | +#endif // __JITCV_H | ... | ... |
sdk/plugins/llvm.cmake
| ... | ... | @@ -21,6 +21,8 @@ if(${BR_WITH_LLVM}) |
| 21 | 21 | # Let's suppose we want to build a JIT compiler with support for binary code: |
| 22 | 22 | llvm_map_components_to_libraries(REQ_LLVM_LIBRARIES jit native) |
| 23 | 23 | |
| 24 | + set(BR_THIRDPARTY_SRC ${BR_THIRDPARTY_SRC} | |
| 25 | + ${CMAKE_SOURCE_DIR}/sdk/plugins/llvm.cpp | |
| 26 | + ${CMAKE_SOURCE_DIR}/sdk/jitcv/jitcv.cpp) | |
| 24 | 27 | set(BR_THIRDPARTY_LIBS ${BR_THIRDPARTY_LIBS} ${REQ_LLVM_LIBRARIES}) |
| 25 | - set(BR_THIRDPARTY_SRC ${BR_THIRDPARTY_SRC} ${CMAKE_SOURCE_DIR}/sdk/plugins/llvm.cpp) | |
| 26 | 28 | endif() | ... | ... |
sdk/plugins/llvm.cpp
| ... | ... | @@ -21,9 +21,11 @@ |
| 21 | 21 | #include <openbr_plugin.h> |
| 22 | 22 | |
| 23 | 23 | #include "core/opencvutils.h" |
| 24 | +#include "jitcv/jitcv.h" | |
| 24 | 25 | |
| 25 | 26 | using namespace br; |
| 26 | 27 | using namespace cv; |
| 28 | +using namespace jitcv; | |
| 27 | 29 | using namespace llvm; |
| 28 | 30 | |
| 29 | 31 | static Module *TheModule = NULL; |
| ... | ... | @@ -32,124 +34,59 @@ static FunctionPassManager *TheFunctionPassManager = NULL; |
| 32 | 34 | static FunctionPassManager *TheExtraFunctionPassManager = NULL; |
| 33 | 35 | static StructType *TheMatrixStruct = NULL; |
| 34 | 36 | |
| 35 | -/*! | |
| 36 | - * \brief LLVM matrix | |
| 37 | - * \author Josh Klontz \cite jklontz | |
| 38 | - * \note Not part of the core SDK | |
| 39 | - */ | |
| 40 | -struct Matrix | |
| 37 | +static QString MatrixToString(const Matrix &m) | |
| 41 | 38 | { |
| 42 | - quint8 *data; /*!< Data */ | |
| 43 | - quint32 channels; /*!< Channels */ | |
| 44 | - quint32 columns; /*!< Columns */ | |
| 45 | - quint32 rows; /*!< Rows */ | |
| 46 | - quint32 frames; /*!< Frames */ | |
| 47 | - quint16 hash; /*!< Bits : 8 | |
| 48 | - Floating : 1 | |
| 49 | - Signed : 1 | |
| 50 | - Single-channel : 1 | |
| 51 | - Single-column : 1 | |
| 52 | - Single-row : 1 | |
| 53 | - Single-frame : 1 */ | |
| 54 | - | |
| 55 | - enum Hash { Bits = 0x00FF, | |
| 56 | - Floating = 0x0100, | |
| 57 | - Signed = 0x0200, | |
| 58 | - SingleChannel = 0x1000, | |
| 59 | - SingleColumn = 0x2000, | |
| 60 | - SingleRow = 0x4000, | |
| 61 | - SingleFrame = 0x8000, | |
| 62 | - u1 = 1, | |
| 63 | - u8 = 8, | |
| 64 | - u16 = 16, | |
| 65 | - u32 = 32, | |
| 66 | - u64 = 64, | |
| 67 | - s8 = 8 + Signed, | |
| 68 | - s16 = 16 + Signed, | |
| 69 | - s32 = 32 + Signed, | |
| 70 | - s64 = 64 + Signed, | |
| 71 | - f16 = 16 + Floating + Signed, | |
| 72 | - f32 = 32 + Floating + Signed, | |
| 73 | - f64 = 64 + Floating + Signed }; | |
| 74 | - | |
| 75 | - Matrix() : data(NULL), channels(0), columns(0), rows(0), frames(0), hash(0) {} | |
| 76 | - | |
| 77 | - Matrix(quint32 _channels, quint32 _columns, quint32 _rows, quint32 _frames, quint16 _hash) | |
| 78 | - : data(NULL), channels(_channels), columns(_columns), rows(_rows), frames(_frames), hash(_hash) | |
| 79 | - { | |
| 80 | - setSingleChannel(channels == 1); | |
| 81 | - setSingleColumn(columns == 1); | |
| 82 | - setSingleRow(rows == 1); | |
| 83 | - setSingleFrame(frames == 1); | |
| 84 | - } | |
| 39 | + return QString("%1%2%3%4%5%6%7").arg(QString::number(m.bits()), (m.isSigned() ? "s" : "u"), (m.isFloating() ? "f" : "i"), | |
| 40 | + QString::number(m.singleChannel()), QString::number(m.singleColumn()), QString::number(m.singleRow()), QString::number(m.singleFrame())); | |
| 41 | +} | |
| 85 | 42 | |
| 86 | - Matrix(const cv::Mat &mat) | |
| 87 | - { | |
| 88 | - if (!mat.isContinuous()) qFatal("Matrix requires continuous data."); | |
| 89 | - channels = mat.channels(); | |
| 90 | - columns = mat.cols; | |
| 91 | - rows = mat.rows; | |
| 92 | - frames = 1; | |
| 93 | - | |
| 94 | - switch (mat.depth()) { | |
| 95 | - case CV_8U: hash = Matrix::u8; break; | |
| 96 | - case CV_8S: hash = Matrix::s8; break; | |
| 97 | - case CV_16U: hash = Matrix::u16; break; | |
| 98 | - case CV_16S: hash = Matrix::s16; break; | |
| 99 | - case CV_32S: hash = Matrix::s32; break; | |
| 100 | - case CV_32F: hash = Matrix::f32; break; | |
| 101 | - case CV_64F: hash = Matrix::f64; break; | |
| 102 | - default: qFatal("Unrecognized matrix depth."); | |
| 103 | - } | |
| 43 | +static Matrix MatrixFromMat(const cv::Mat &mat) | |
| 44 | +{ | |
| 45 | + Matrix m; | |
| 46 | + | |
| 47 | + if (!mat.isContinuous()) qFatal("Matrix requires continuous data."); | |
| 48 | + m.channels = mat.channels(); | |
| 49 | + m.columns = mat.cols; | |
| 50 | + m.rows = mat.rows; | |
| 51 | + m.frames = 1; | |
| 52 | + | |
| 53 | + switch (mat.depth()) { | |
| 54 | + case CV_8U: m.hash = Matrix::u8; break; | |
| 55 | + case CV_8S: m.hash = Matrix::s8; break; | |
| 56 | + case CV_16U: m.hash = Matrix::u16; break; | |
| 57 | + case CV_16S: m.hash = Matrix::s16; break; | |
| 58 | + case CV_32S: m.hash = Matrix::s32; break; | |
| 59 | + case CV_32F: m.hash = Matrix::f32; break; | |
| 60 | + case CV_64F: m.hash = Matrix::f64; break; | |
| 61 | + default: qFatal("Unrecognized matrix depth."); | |
| 62 | + } | |
| 63 | + | |
| 64 | + m.data = mat.data; | |
| 65 | + return m; | |
| 66 | +} | |
| 104 | 67 | |
| 105 | - data = mat.data; | |
| 106 | - } | |
| 107 | - | |
| 108 | - void copyHeader(const Matrix &other) { channels = other.channels; columns = other.columns; rows = other.rows; frames = other.frames; hash = other.hash; } | |
| 109 | - void allocate() { deallocate(); data = new quint8[bytes()]; } | |
| 110 | - void allocate(Mat &m) { deallocate(); m = Mat(rows, columns, CV_MAKETYPE(getOpenCVDepth(), channels)); data = m.data; } | |
| 111 | - void deallocate() { delete[] data; data = NULL; } | |
| 112 | - | |
| 113 | - inline int bits() const { return hash & Bits; } | |
| 114 | - inline void setBits(int bits) { hash &= ~Bits; hash |= bits & Bits; } | |
| 115 | - inline bool isFloating() const { return hash & Floating; } | |
| 116 | - inline void setFloating(bool isFloating) { isFloating ? setSigned(true), hash |= Floating : hash &= ~Floating; } | |
| 117 | - inline bool isSigned() const { return hash & Signed; } | |
| 118 | - inline void setSigned(bool isSigned) { isSigned ? hash |= Signed : hash &= ~Signed; } | |
| 119 | - inline int type() const { return hash & (Bits + Floating + Signed); } | |
| 120 | - inline void setType(int type) { hash &= ~(Bits + Floating + Signed); hash |= type & (Bits + Floating + Signed); } | |
| 121 | - inline bool singleChannel() const { return hash & SingleChannel; } | |
| 122 | - inline void setSingleChannel(bool singleChannel) { singleChannel ? hash |= SingleChannel : hash &= ~SingleChannel; } | |
| 123 | - inline bool singleColumn() const { return hash & SingleColumn; } | |
| 124 | - inline void setSingleColumn(bool singleColumn) { singleColumn ? hash |= SingleColumn : hash &= ~SingleColumn; } | |
| 125 | - inline bool singleRow() const { return hash & SingleRow; } | |
| 126 | - inline void setSingleRow(bool singleRow) { singleRow ? hash |= SingleRow : hash &= ~SingleRow; } | |
| 127 | - inline bool singleFrame() const { return hash & SingleFrame; } | |
| 128 | - inline void setSingleFrame(bool singleFrame) { singleFrame ? hash |= SingleFrame : hash &= ~SingleFrame; } | |
| 129 | - inline quint32 elements() const { return channels * columns * rows * frames; } | |
| 130 | - inline quint32 bytes() const { return bits() / 8 * elements(); } | |
| 131 | - | |
| 132 | - QString toString() const { return QString("%1%2%3%4%5%6%7").arg(QString::number(bits()), (isSigned() ? "s" : "u"), (isFloating() ? "f" : "i"), | |
| 133 | - QString::number(singleChannel()), QString::number(singleColumn()), QString::number(singleRow()), QString::number(singleFrame())); } | |
| 134 | - | |
| 135 | - int getOpenCVDepth() const { | |
| 136 | - switch (type()) { | |
| 137 | - case u8: return CV_8U; | |
| 138 | - case s8: return CV_8S; | |
| 139 | - case u16: return CV_16U; | |
| 140 | - case s16: return CV_16S; | |
| 141 | - case s32: return CV_32S; | |
| 142 | - case f32: return CV_32F; | |
| 143 | - case f64: return CV_64F; | |
| 144 | - default: qFatal("OpenCV does not support Matrix format: %s", qPrintable(toString())); | |
| 145 | - } | |
| 146 | - return -1; | |
| 147 | - } | |
| 148 | -}; | |
| 68 | +static void AllocateMatrixFromMat(Matrix &m, cv::Mat &mat) | |
| 69 | +{ | |
| 70 | + int cvType = -1; | |
| 71 | + switch (m.type()) { | |
| 72 | + case Matrix::u8: cvType = CV_8U; break; | |
| 73 | + case Matrix::s8: cvType = CV_8S; break; | |
| 74 | + case Matrix::u16: cvType = CV_16U; break; | |
| 75 | + case Matrix::s16: cvType = CV_16S; break; | |
| 76 | + case Matrix::s32: cvType = CV_32S; break; | |
| 77 | + case Matrix::f32: cvType = CV_32F; break; | |
| 78 | + case Matrix::f64: cvType = CV_64F; break; | |
| 79 | + default: qFatal("OpenCV does not support Matrix format: %s", qPrintable(MatrixToString(m))); | |
| 80 | + } | |
| 81 | + | |
| 82 | + m.deallocate(); | |
| 83 | + mat = Mat(m.rows, m.columns, CV_MAKETYPE(cvType, m.channels)); | |
| 84 | + m.data = mat.data; | |
| 85 | +} | |
| 149 | 86 | |
| 150 | 87 | QDebug operator<<(QDebug dbg, const Matrix &m) |
| 151 | 88 | { |
| 152 | - dbg.nospace() << m.toString(); | |
| 89 | + dbg.nospace() << MatrixToString(m); | |
| 153 | 90 | return dbg; |
| 154 | 91 | } |
| 155 | 92 | |
| ... | ... | @@ -307,10 +244,10 @@ namespace br |
| 307 | 244 | { |
| 308 | 245 | |
| 309 | 246 | /*! |
| 310 | - * \brief LLVM Kernel | |
| 247 | + * \brief LLVM Unary Kernel | |
| 311 | 248 | * \author Josh Klontz \cite jklontz |
| 312 | 249 | */ |
| 313 | -class Kernel : public UntrainableMetaTransform | |
| 250 | +class UnaryKernel : public UntrainableMetaTransform | |
| 314 | 251 | { |
| 315 | 252 | Q_OBJECT |
| 316 | 253 | |
| ... | ... | @@ -319,17 +256,18 @@ class Kernel : public UntrainableMetaTransform |
| 319 | 256 | quint16 hash; |
| 320 | 257 | |
| 321 | 258 | public: |
| 322 | - Kernel() : kernel(NULL), hash(0) {} | |
| 259 | + UnaryKernel() : kernel(NULL), hash(0) {} | |
| 323 | 260 | virtual int preallocate(const Matrix &src, Matrix &dst) const = 0; /*!< Preallocate destintation matrix based on source matrix. */ |
| 324 | 261 | virtual void build(const MatrixBuilder &src, const MatrixBuilder &dst, PHINode *i) const = 0; /*!< Build the kernel. */ |
| 325 | 262 | |
| 326 | 263 | private: |
| 327 | - QString mangledName(const Matrix &m) const { | |
| 264 | + QString mangledName(const Matrix &src) const | |
| 265 | + { | |
| 328 | 266 | static QHash<QString, int> argsLUT; |
| 329 | 267 | const QString args = arguments().join(","); |
| 330 | 268 | if (!argsLUT.contains(args)) argsLUT.insert(args, argsLUT.size()); |
| 331 | 269 | int uid = argsLUT.value(args); |
| 332 | - return "jitcv_" + name().remove("Transform") + QString::number(uid) + "_" + m.toString(); | |
| 270 | + return "jitcv_" + name().remove("Transform") + (args.isEmpty() ? QString() : QString::number(uid)) + "_" + MatrixToString(src); | |
| 333 | 271 | } |
| 334 | 272 | |
| 335 | 273 | Function *compile(const Matrix &m) const |
| ... | ... | @@ -370,32 +308,125 @@ private: |
| 370 | 308 | |
| 371 | 309 | void project(const Template &src, Template &dst) const |
| 372 | 310 | { |
| 373 | - const Matrix m(src.m()); | |
| 311 | + const Matrix m(MatrixFromMat(src)); | |
| 374 | 312 | Matrix n; |
| 375 | 313 | const int size = preallocate(m, n); |
| 376 | - n.allocate(dst.m()); | |
| 314 | + AllocateMatrixFromMat(n, dst); | |
| 315 | + invoke(m, n, size); | |
| 316 | + } | |
| 317 | + | |
| 318 | + void invoke(const Matrix &src, Matrix &dst, int size) const | |
| 319 | + { | |
| 320 | + if (src.hash != hash) { | |
| 321 | + static QMutex compilerLock; | |
| 322 | + QMutexLocker locker(&compilerLock); | |
| 323 | + | |
| 324 | + if (src.hash != hash) { | |
| 325 | + const QString functionName = mangledName(src); | |
| 326 | + | |
| 327 | + Function *function = TheModule->getFunction(qPrintable(functionName)); | |
| 328 | + if (function == NULL) { | |
| 329 | + function = compile(src); | |
| 330 | + while (TheFunctionPassManager->run(*function)); | |
| 331 | + TheExtraFunctionPassManager->run(*function); | |
| 332 | + function = TheModule->getFunction(qPrintable(functionName)); | |
| 333 | + } | |
| 334 | + | |
| 335 | + const_cast<UnaryKernel*>(this)->kernel = (kernel_t)TheExecutionEngine->getPointerToFunction(function); | |
| 336 | + const_cast<UnaryKernel*>(this)->hash = src.hash; | |
| 337 | + } | |
| 338 | + } | |
| 339 | + | |
| 340 | + kernel(&src, &dst, size); | |
| 341 | + } | |
| 342 | +}; | |
| 343 | + | |
| 344 | +/*! | |
| 345 | + * \brief LLVM Binary Kernel | |
| 346 | + * \author Josh Klontz \cite jklontz | |
| 347 | + */ | |
| 348 | +class BinaryKernel: public UntrainableMetaTransform | |
| 349 | +{ | |
| 350 | + Q_OBJECT | |
| 351 | + | |
| 352 | + typedef void (*kernel_t)(const Matrix*, const Matrix*, Matrix*, quint32); | |
| 353 | + kernel_t kernel; | |
| 354 | + quint16 hashA, hashB; | |
| 355 | + | |
| 356 | +public: | |
| 357 | + BinaryKernel() : kernel(NULL), hashA(0), hashB(0) {} | |
| 358 | + virtual int preallocate(const Matrix &srcA, const Matrix &srcB, Matrix &dst) const = 0; /*!< Preallocate destintation matrix based on source matrix. */ | |
| 359 | + virtual void build(const MatrixBuilder &srcA, const MatrixBuilder &srcB, const MatrixBuilder &dst, PHINode *i) const = 0; /*!< Build the kernel. */ | |
| 360 | + | |
| 361 | +private: | |
| 362 | + QString mangledName(const Matrix &srcA, const Matrix &srcB) const | |
| 363 | + { | |
| 364 | + return "jitcv_" + name().remove("Transform") + "_" + MatrixToString(srcA) + "_" + MatrixToString(srcB); | |
| 365 | + } | |
| 366 | + | |
| 367 | + Function *compile(const Matrix &m, const Matrix &n) const | |
| 368 | + { | |
| 369 | + Constant *c = TheModule->getOrInsertFunction(qPrintable(mangledName(m, n)), | |
| 370 | + Type::getVoidTy(getGlobalContext()), | |
| 371 | + PointerType::getUnqual(TheMatrixStruct), | |
| 372 | + PointerType::getUnqual(TheMatrixStruct), | |
| 373 | + PointerType::getUnqual(TheMatrixStruct), | |
| 374 | + Type::getInt32Ty(getGlobalContext()), | |
| 375 | + NULL); | |
| 376 | + | |
| 377 | + Function *function = cast<Function>(c); | |
| 378 | + function->setCallingConv(CallingConv::C); | |
| 377 | 379 | |
| 378 | - if (m.hash != hash) { | |
| 380 | + Function::arg_iterator args = function->arg_begin(); | |
| 381 | + Value *srcA = args++; | |
| 382 | + srcA->setName("srcA"); | |
| 383 | + Value *srcB = args++; | |
| 384 | + srcB->setName("srcB"); | |
| 385 | + Value *dst = args++; | |
| 386 | + dst->setName("dst"); | |
| 387 | + Value *len = args++; | |
| 388 | + len->setName("len"); | |
| 389 | + | |
| 390 | + BasicBlock *entry = BasicBlock::Create(getGlobalContext(), "entry", function); | |
| 391 | + IRBuilder<> builder(entry); | |
| 392 | + | |
| 393 | + BasicBlock *kernel; | |
| 394 | + PHINode *i = MatrixBuilder::beginLoop(builder, function, entry, &kernel, "i"); | |
| 395 | + | |
| 396 | + Matrix o; | |
| 397 | + preallocate(m, n, o); | |
| 398 | + build(MatrixBuilder(m, srcA, &builder, function, "srcA"), MatrixBuilder(n, srcB, &builder, function, "srcB"), MatrixBuilder(o, dst, &builder, function, "dst"), i); | |
| 399 | + | |
| 400 | + MatrixBuilder::endLoop(builder, function, kernel, i, len, "i"); | |
| 401 | + | |
| 402 | + builder.CreateRetVoid(); | |
| 403 | + return function; | |
| 404 | + } | |
| 405 | + | |
| 406 | + void invoke(const Matrix &srcA, const Matrix &srcB, Matrix &dst, int size) const | |
| 407 | + { | |
| 408 | + if ((srcA.hash != hashA) || (srcB.hash != hashB)) { | |
| 379 | 409 | static QMutex compilerLock; |
| 380 | 410 | QMutexLocker locker(&compilerLock); |
| 381 | 411 | |
| 382 | - if (m.hash != hash) { | |
| 383 | - const QString functionName = mangledName(m); | |
| 412 | + if ((srcA.hash != hashA) || (srcB.hash != hashB)) { | |
| 413 | + const QString functionName = mangledName(srcA, srcB); | |
| 384 | 414 | |
| 385 | 415 | Function *function = TheModule->getFunction(qPrintable(functionName)); |
| 386 | 416 | if (function == NULL) { |
| 387 | - function = compile(m); | |
| 417 | + function = compile(srcA, srcB); | |
| 388 | 418 | while (TheFunctionPassManager->run(*function)); |
| 389 | 419 | TheExtraFunctionPassManager->run(*function); |
| 390 | 420 | function = TheModule->getFunction(qPrintable(functionName)); |
| 391 | 421 | } |
| 392 | 422 | |
| 393 | - const_cast<Kernel*>(this)->kernel = (kernel_t)TheExecutionEngine->getPointerToFunction(function); | |
| 394 | - const_cast<Kernel*>(this)->hash = m.hash; | |
| 423 | + const_cast<BinaryKernel*>(this)->kernel = (kernel_t)TheExecutionEngine->getPointerToFunction(function); | |
| 424 | + const_cast<BinaryKernel*>(this)->hashA = srcA.hash; | |
| 425 | + const_cast<BinaryKernel*>(this)->hashB = srcB.hash; | |
| 395 | 426 | } |
| 396 | 427 | } |
| 397 | 428 | |
| 398 | - kernel(&m, &n, size); | |
| 429 | + kernel(&srcA, &srcB, &dst, size); | |
| 399 | 430 | } |
| 400 | 431 | }; |
| 401 | 432 | |
| ... | ... | @@ -403,7 +434,7 @@ private: |
| 403 | 434 | * \brief LLVM Stitchable Kernel |
| 404 | 435 | * \author Josh Klontz \cite jklontz |
| 405 | 436 | */ |
| 406 | -class StitchableKernel : public Kernel | |
| 437 | +class StitchableKernel : public UnaryKernel | |
| 407 | 438 | { |
| 408 | 439 | Q_OBJECT |
| 409 | 440 | |
| ... | ... | @@ -430,7 +461,7 @@ private: |
| 430 | 461 | * \brief LLVM stitch transform |
| 431 | 462 | * \author Josh Klontz \cite jklontz |
| 432 | 463 | */ |
| 433 | -class stitchTransform : public Kernel | |
| 464 | +class stitchTransform : public UnaryKernel | |
| 434 | 465 | { |
| 435 | 466 | Q_OBJECT |
| 436 | 467 | Q_PROPERTY(QList<br::Transform*> kernels READ get_kernels WRITE set_kernels RESET reset_kernels STORED false) |
| ... | ... | @@ -447,7 +478,7 @@ class stitchTransform : public Kernel |
| 447 | 478 | { |
| 448 | 479 | Matrix tmp = src; |
| 449 | 480 | foreach (const Transform *kernel, kernels) { |
| 450 | - static_cast<const Kernel*>(kernel)->preallocate(tmp, dst); | |
| 481 | + static_cast<const UnaryKernel*>(kernel)->preallocate(tmp, dst); | |
| 451 | 482 | tmp = dst; |
| 452 | 483 | } |
| 453 | 484 | return dst.elements(); |
| ... | ... | @@ -459,7 +490,7 @@ class stitchTransform : public Kernel |
| 459 | 490 | MatrixBuilder dst(dst_); |
| 460 | 491 | Value *val = src.load(i); |
| 461 | 492 | foreach (Transform *transform, kernels) { |
| 462 | - static_cast<Kernel*>(transform)->preallocate(src, dst); | |
| 493 | + static_cast<UnaryKernel*>(transform)->preallocate(src, dst); | |
| 463 | 494 | val = static_cast<StitchableKernel*>(transform)->stitch(src, dst, val); |
| 464 | 495 | src.copyHeader(dst); |
| 465 | 496 | src.m = dst.m; |
| ... | ... | @@ -532,7 +563,7 @@ BR_REGISTER(Transform, powTransform) |
| 532 | 563 | * \brief LLVM sum transform |
| 533 | 564 | * \author Josh Klontz \cite jklontz |
| 534 | 565 | */ |
| 535 | -class sumTransform : public Kernel | |
| 566 | +class sumTransform : public UnaryKernel | |
| 536 | 567 | { |
| 537 | 568 | Q_OBJECT |
| 538 | 569 | Q_PROPERTY(bool channels READ get_channels WRITE set_channels RESET reset_channels STORED false) |
| ... | ... | @@ -823,24 +854,24 @@ class LLVMInitializer : public Initializer |
| 823 | 854 | Type::getInt16Ty(getGlobalContext()), // hash |
| 824 | 855 | NULL); |
| 825 | 856 | |
| 826 | -// QSharedPointer<Transform> kernel(Transform::make("abs", NULL)); | |
| 857 | + QSharedPointer<Transform> kernel(Transform::make("abs", NULL)); | |
| 827 | 858 | |
| 828 | -// Template src, dst; | |
| 829 | -// src.m() = (Mat_<qint8>(2,2) << -1, -2, 3, 4); | |
| 830 | -// kernel->project(src, dst); | |
| 831 | -// qDebug() << dst.m(); | |
| 859 | + Template src, dst; | |
| 860 | + src.m() = (Mat_<qint8>(2,2) << -1, -2, 3, 4); | |
| 861 | + kernel->project(src, dst); | |
| 862 | + qDebug() << dst.m(); | |
| 832 | 863 | |
| 833 | -// src.m() = (Mat_<qint32>(2,2) << -1, -3, 9, 27); | |
| 834 | -// kernel->project(src, dst); | |
| 835 | -// qDebug() << dst.m(); | |
| 864 | + src.m() = (Mat_<qint32>(2,2) << -1, -3, 9, 27); | |
| 865 | + kernel->project(src, dst); | |
| 866 | + qDebug() << dst.m(); | |
| 836 | 867 | |
| 837 | -// src.m() = (Mat_<float>(2,2) << -1.5, -2.5, 3.5, 4.5); | |
| 838 | -// kernel->project(src, dst); | |
| 839 | -// qDebug() << dst.m(); | |
| 868 | + src.m() = (Mat_<float>(2,2) << -1.5, -2.5, 3.5, 4.5); | |
| 869 | + kernel->project(src, dst); | |
| 870 | + qDebug() << dst.m(); | |
| 840 | 871 | |
| 841 | -// src.m() = (Mat_<double>(2,2) << 1.75, 2.75, -3.75, -4.75); | |
| 842 | -// kernel->project(src, dst); | |
| 843 | -// qDebug() << dst.m(); | |
| 872 | + src.m() = (Mat_<double>(2,2) << 1.75, 2.75, -3.75, -4.75); | |
| 873 | + kernel->project(src, dst); | |
| 874 | + qDebug() << dst.m(); | |
| 844 | 875 | } |
| 845 | 876 | |
| 846 | 877 | void finalize() const | ... | ... |
sdk/plugins/plugins.cmake