colin
/
CuPBoP
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge branch 'cupbop:master' into master

This commit is contained in:
Jun Chen 2022-05-24 21:16:37 -04:00 committed by GitHub
parent 01fcd6e0cc 197abc867d
commit 21f298524e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
27 changed files with 1246 additions and 1384 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
README.md
View File

@ -27,10 +27,10 @@ Currently, CuPBoP support serveral CPU backends, including x86, AArch64, and RIS
export CuPBoP_PATH=`pwd` export CuPBoP_PATH=`pwd`
export LD_LIBRARY_PATH=$CuPBoP_PATH/build/runtime:$CuPBoP_PATH/build/runtime/threadPool:$LD_LIBRARY_PATH export LD_LIBRARY_PATH=$CuPBoP_PATH/build/runtime:$CuPBoP_PATH/build/runtime/threadPool:$LD_LIBRARY_PATH
``` ```
If you are using boson, you can pre-installed llvm 10.0.0
LLVM_PATH=/opt/llvm-10.0.0 If you are using boson, you can pre-installed llvm 10.0.0\
export PATH=$LLVM_PATH/bin:$PATH `LLVM_PATH=/opt/llvm-10.0.0`\
`export PATH=$LLVM_PATH/bin:$PATH`
2. As CuPBoP relies on CUDA structures, we need to download the CUDA header file 2. As CuPBoP relies on CUDA structures, we need to download the CUDA header file

2
compilation/KernelTranslation/lib/init.cpp
View File

@ -396,7 +396,7 @@ void init_block(llvm::Module *M, std::ofstream &fout) {
replace_asm_call(M); replace_asm_call(M);
// replace dynamic shared memory // replace dynamic shared memory
auto dynamic_shared_memory_addr = auto dynamic_shared_memory_addr =
M->getGlobalVariable("dynamic_shared_memory"); M->getGlobalVariable("dynamic_shared_memory");
if (dynamic_shared_memory_addr) { if (dynamic_shared_memory_addr) {
replace_dynamic_shared_memory(M); replace_dynamic_shared_memory(M);
} }

3
compilation/KernelTranslation/lib/insert_warp_loop.cpp
View File

@ -272,13 +272,12 @@ void AddContextSaveRestore(llvm::Instruction *instruction,
std::vector<Instruction *> uses; std::vector<Instruction *> uses;
Function *f2 = instruction->getParent()->getParent(); Function *f2 = instruction->getParent()->getParent();
for (Instruction::use_iterator ui = instruction->use_begin(), for (Instruction::use_iterator ui = instruction->use_begin(),
ue = instruction->use_end(); ue = instruction->use_end();
ui != ue; ++ui) { ui != ue; ++ui) {
llvm::Instruction *user = cast<Instruction>(ui->getUser()); llvm::Instruction *user = cast<Instruction>(ui->getUser());
Function *f1 = user->getParent()->getParent(); Function *f1 = user->getParent()->getParent();
if(f2->getName() != f1->getName()) { if (f2->getName() != f1->getName()) {
continue; continue;
} }
if (user == NULL) if (user == NULL)

12
compilation/KernelTranslation/lib/memory_hierarchy.cpp
View File

@ -89,11 +89,12 @@ void mem_share2global(llvm::Module *M) {
} else if (element_type->isStructTy()) { } else if (element_type->isStructTy()) {
auto undef = llvm::UndefValue::get(element_type); auto undef = llvm::UndefValue::get(element_type);
llvm::GlobalVariable *global_memory = new llvm::GlobalVariable( llvm::GlobalVariable *global_memory = new llvm::GlobalVariable(
*M, element_type, false, llvm::GlobalValue::ExternalLinkage, undef, *M, element_type, false, llvm::GlobalValue::ExternalLinkage,
new_name, NULL, llvm::GlobalValue::GeneralDynamicTLSModel, 0, undef, new_name, NULL,
false); llvm::GlobalValue::GeneralDynamicTLSModel, 0, false);
global_memory->setDSOLocal(true); global_memory->setDSOLocal(true);
Comdat * comdat = M->getOrInsertComdat(StringRef(share_memory->getName())); Comdat *comdat =
M->getOrInsertComdat(StringRef(share_memory->getName()));
comdat->setSelectionKind(Comdat::SelectionKind::Any); comdat->setSelectionKind(Comdat::SelectionKind::Any);
global_memory->setComdat(comdat); global_memory->setComdat(comdat);
global_memory->setLinkage(llvm::GlobalValue::LinkOnceODRLinkage); global_memory->setLinkage(llvm::GlobalValue::LinkOnceODRLinkage);
@ -101,8 +102,7 @@ void mem_share2global(llvm::Module *M) {
global_memory->setAlignment(share_memory->getAlignment()); global_memory->setAlignment(share_memory->getAlignment());
corresponding_global_memory.insert( corresponding_global_memory.insert(
std::pair<GlobalVariable *, GlobalVariable *>(share_memory, std::pair<GlobalVariable *, GlobalVariable *>(share_memory,
global_memory)); global_memory));
} else { } else {
assert(0 && "The required Share Memory Type is not supported\n"); assert(0 && "The required Share Memory Type is not supported\n");

34
examples/dwt2d/common.h Executable file → Normal file
View File

@ -27,9 +27,9 @@
#ifndef _COMMON_H #ifndef _COMMON_H
#define _COMMON_H #define _COMMON_H
//24-bit multiplication is faster on G80, // 24-bit multiplication is faster on G80,
//but we must be sure to multiply integers // but we must be sure to multiply integers
//only within [-8M, 8M - 1] range // only within [-8M, 8M - 1] range
#define IMUL(a, b) __mul24(a, b) #define IMUL(a, b) __mul24(a, b)
////cuda timing macros ////cuda timing macros
@ -42,21 +42,23 @@
// cudaEventSynchronize(cstop); \ // cudaEventSynchronize(cstop); \
// cudaEventElapsedTime(&elapsedTime, cstart, cstop) // cudaEventElapsedTime(&elapsedTime, cstart, cstop)
//divide and round up macro // divide and round up macro
#define DIVANDRND(a, b) ((((a) % (b)) != 0) ? ((a) / (b) + 1) : ((a) / (b))) #define DIVANDRND(a, b) ((((a) % (b)) != 0) ? ((a) / (b) + 1) : ((a) / (b)))
# define cudaCheckError( msg ) { \ #define cudaCheckError(msg) \
cudaError_t err = cudaGetLastError(); \ { \
if( cudaSuccess != err) { \ cudaError_t err = cudaGetLastError(); \
fprintf(stderr, "%s: %i: %s: %s.\n", \ if (cudaSuccess != err) { \
__FILE__, __LINE__, msg, cudaGetErrorString( err) ); \ fprintf(stderr, "%s: %i: %s: %s.\n", __FILE__, __LINE__, msg, \
exit(-1); \ cudaGetErrorString(err)); \
} } exit(-1); \
} \
# define cudaCheckAsyncError( msg ) { \ }
cudaThreadSynchronize(); \
cudaCheckError( msg ); \
}
#define cudaCheckAsyncError(msg) \
{ \
cudaThreadSynchronize(); \
cudaCheckError(msg); \
}
#endif #endif

9
examples/dwt2d/components.h Executable file → Normal file
View File

@ -28,11 +28,12 @@
#define _COMPONENTS_H #define _COMPONENTS_H
/* Separate compoents of source 8bit RGB image */ /* Separate compoents of source 8bit RGB image */
template<typename T> template <typename T>
void rgbToComponents(T *d_r, T *d_g, T *d_b, unsigned char * src, int width, int height); void rgbToComponents(T *d_r, T *d_g, T *d_b, unsigned char *src, int width,
int height);
/* Copy a 8bit source image data into a color compoment of type T */ /* Copy a 8bit source image data into a color compoment of type T */
template<typename T> template <typename T>
void bwToComponent(T *d_c, unsigned char * src, int width, int height); void bwToComponent(T *d_c, unsigned char *src, int width, int height);
#endif #endif

17
examples/dwt2d/dwt.h Executable file → Normal file
View File

@ -27,14 +27,15 @@
#ifndef _DWT_H #ifndef _DWT_H
#define _DWT_H #define _DWT_H
template<typename T> template <typename T>
int nStage2dDWT(T *in, T *out, T * backup, int pixWidth, int pixHeight, int stages, bool forward); int nStage2dDWT(T *in, T *out, T *backup, int pixWidth, int pixHeight,
int stages, bool forward);
template<typename T> template <typename T>
int writeNStage2DDWT(T *component_cuda, int width, int height, int writeNStage2DDWT(T *component_cuda, int width, int height, int stages,
int stages, const char * filename, const char * suffix); const char *filename, const char *suffix);
template<typename T> template <typename T>
int writeLinear(T *component_cuda, int width, int height, int writeLinear(T *component_cuda, int width, int height, const char *filename,
const char * filename, const char * suffix); const char *suffix);
#endif #endif

369
examples/dwt2d/dwt_cuda/common.h Executable file → Normal file
View File

@ -29,233 +29,204 @@
/// POSSIBILITY OF SUCH DAMAGE. /// POSSIBILITY OF SUCH DAMAGE.
/// ///
#ifndef DWT_COMMON_H #ifndef DWT_COMMON_H
#define DWT_COMMON_H #define DWT_COMMON_H
#include <cstdio>
#include <algorithm> #include <algorithm>
#include <cstdio>
#include <vector> #include <vector>
// compile time minimum macro // compile time minimum macro
#define CTMIN(a,b) (((a) < (b)) ? (a) : (b)) #define CTMIN(a, b) (((a) < (b)) ? (a) : (b))
// performance testing macros // performance testing macros
#if defined(GPU_DWT_TESTING) #if defined(GPU_DWT_TESTING)
#define PERF_BEGIN \ #define PERF_BEGIN \
{ \ { \
dwt_cuda::CudaDWTTester PERF_TESTER; \ dwt_cuda::CudaDWTTester PERF_TESTER; \
for(int PERF_N = PERF_TESTER.getNumIterations(); PERF_N--; ) \ for (int PERF_N = PERF_TESTER.getNumIterations(); PERF_N--;) { \
{ \
PERF_TESTER.beginTestIteration(); PERF_TESTER.beginTestIteration();
#define PERF_END(PERF_NAME, PERF_W, PERF_H) \ #define PERF_END(PERF_NAME, PERF_W, PERF_H) \
PERF_TESTER.endTestIteration(); \ PERF_TESTER.endTestIteration(); \
} \ } \
PERF_TESTER.showPerformance(PERF_NAME, PERF_W, PERF_H); \ PERF_TESTER.showPerformance(PERF_NAME, PERF_W, PERF_H); \
} }
#else // GPU_DWT_TESTING #else // GPU_DWT_TESTING
#define PERF_BEGIN #define PERF_BEGIN
#define PERF_END(PERF_NAME, PERF_W, PERF_H) #define PERF_END(PERF_NAME, PERF_W, PERF_H)
#endif // GPU_DWT_TESTING #endif // GPU_DWT_TESTING
namespace dwt_cuda { namespace dwt_cuda {
/// Divide and round up.
template <typename T>
__device__ __host__ inline T divRndUp(const T &n, const T &d) {
return (n / d) + ((n % d) ? 1 : 0);
}
/// Divide and round up. // 9/7 forward DWT lifting schema coefficients
template <typename T> const float f97Predict1 = -1.586134342; ///< forward 9/7 predict 1
__device__ __host__ inline T divRndUp(const T & n, const T & d) { const float f97Update1 = -0.05298011854; ///< forward 9/7 update 1
return (n / d) + ((n % d) ? 1 : 0); const float f97Predict2 = 0.8829110762; ///< forward 9/7 predict 2
const float f97Update2 = 0.4435068522; ///< forward 9/7 update 2
// 9/7 reverse DWT lifting schema coefficients
const float r97update2 = -f97Update2; ///< undo 9/7 update 2
const float r97predict2 = -f97Predict2; ///< undo 9/7 predict 2
const float r97update1 = -f97Update1; ///< undo 9/7 update 1
const float r97Predict1 = -f97Predict1; ///< undo 9/7 predict 1
// FDWT 9/7 scaling coefficients
const float scale97Mul = 1.23017410491400f;
const float scale97Div = 1.0 / scale97Mul;
// 5/3 forward DWT lifting schema coefficients
const float forward53Predict = -0.5f; /// forward 5/3 predict
const float forward53Update = 0.25f; /// forward 5/3 update
// 5/3 forward DWT lifting schema coefficients
const float reverse53Update = -forward53Update; /// undo 5/3 update
const float reverse53Predict = -forward53Predict; /// undo 5/3 predict
/// Functor which adds scaled sum of neighbors to given central pixel.
struct AddScaledSum {
const float scale; // scale of neighbors
__device__ AddScaledSum(const float scale) : scale(scale) {}
__device__ void operator()(const float p, float &c, const float n) const {
// if(threadIdx.x == 0) {
// printf("scale %f, p %f c %f n %f , result: %f\n", scale, p, c, n,
// scale * (p + n) );
// }
c += scale * (p + n);
}
};
/// Returns index ranging from 0 to num threads, such that first half
/// of threads get even indices and others get odd indices. Each thread
/// gets different index.
/// Example: (for 8 threads) threadIdx.x: 0 1 2 3 4 5 6 7
/// parityIdx: 0 2 4 6 1 3 5 7
/// @tparam THREADS total count of participating threads
/// @return parity-separated index of thread
template <int THREADS> __device__ inline int parityIdx() {
return (threadIdx.x * 2) - (THREADS - 1) * (threadIdx.x / (THREADS / 2));
}
/// size of shared memory
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200)
const int SHM_SIZE = 48 * 1024;
#else
const int SHM_SIZE = 16 * 1024;
#endif
/// Perrformance and return code tester.
class CudaDWTTester {
private:
static bool testRunning; ///< true if any test is currently running
cudaEvent_t beginEvent; ///< begin CUDA event
cudaEvent_t endEvent; ///< end CUDA event
std::vector<float> times; ///< collected times
const bool disabled; ///< true if this object is disabled
public:
/// Checks CUDA related error.
/// @param status return code to be checked
/// @param message message to be shown if there was an error
/// @return true if there was no error, false otherwise
static bool check(const cudaError_t &status, const char *message) {
#if defined(GPU_DWT_TESTING)
if ((!testRunning) && status != cudaSuccess) {
const char *errorString = cudaGetErrorString(status);
fprintf(stderr, "CUDA ERROR: '%s': %s\n", message, errorString);
fflush(stderr);
return false;
}
#endif // GPU_DWT_TESTING
return true;
} }
/// Checks last kernel call for errors.
// 9/7 forward DWT lifting schema coefficients /// @param message description of the kernel call
const float f97Predict1 = -1.586134342; ///< forward 9/7 predict 1 /// @return true if there was no error, false otherwise
const float f97Update1 = -0.05298011854; ///< forward 9/7 update 1 static bool checkLastKernelCall(const char *message) {
const float f97Predict2 = 0.8829110762; ///< forward 9/7 predict 2 #if defined(GPU_DWT_TESTING)
const float f97Update2 = 0.4435068522; ///< forward 9/7 update 2 return testRunning ? true : check(cudaThreadSynchronize(), message);
#else // GPU_DWT_TESTING
return true;
// 9/7 reverse DWT lifting schema coefficients #endif // GPU_DWT_TESTING
const float r97update2 = -f97Update2; ///< undo 9/7 update 2
const float r97predict2 = -f97Predict2; ///< undo 9/7 predict 2
const float r97update1 = -f97Update1; ///< undo 9/7 update 1
const float r97Predict1 = -f97Predict1; ///< undo 9/7 predict 1
// FDWT 9/7 scaling coefficients
const float scale97Mul = 1.23017410491400f;
const float scale97Div = 1.0 / scale97Mul;
// 5/3 forward DWT lifting schema coefficients
const float forward53Predict = -0.5f; /// forward 5/3 predict
const float forward53Update = 0.25f; /// forward 5/3 update
// 5/3 forward DWT lifting schema coefficients
const float reverse53Update = -forward53Update; /// undo 5/3 update
const float reverse53Predict = -forward53Predict; /// undo 5/3 predict
/// Functor which adds scaled sum of neighbors to given central pixel.
struct AddScaledSum {
const float scale; // scale of neighbors
__device__ AddScaledSum(const float scale) : scale(scale) {}
__device__ void operator()(const float p, float & c, const float n) const {
// if(threadIdx.x == 0) {
// printf("scale %f, p %f c %f n %f , result: %f\n", scale, p, c, n, scale * (p + n) );
// }
c += scale * (p + n);
}
};
/// Returns index ranging from 0 to num threads, such that first half
/// of threads get even indices and others get odd indices. Each thread
/// gets different index.
/// Example: (for 8 threads) threadIdx.x: 0 1 2 3 4 5 6 7
/// parityIdx: 0 2 4 6 1 3 5 7
/// @tparam THREADS total count of participating threads
/// @return parity-separated index of thread
template <int THREADS>
__device__ inline int parityIdx() {
return (threadIdx.x * 2) - (THREADS - 1) * (threadIdx.x / (THREADS / 2));
} }
/// Initializes DWT tester for time measurement
CudaDWTTester() : disabled(testRunning) {}
/// Gets rpefered number of iterations
int getNumIterations() { return disabled ? 1 : 31; }
/// size of shared memory /// Starts one test iteration.
#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 200) void beginTestIteration() {
const int SHM_SIZE = 48 * 1024; if (!disabled) {
#else cudaEventCreate(&beginEvent);
const int SHM_SIZE = 16 * 1024; cudaEventCreate(&endEvent);
#endif cudaEventRecord(beginEvent, 0);
testRunning = true;
/// Perrformance and return code tester.
class CudaDWTTester {
private:
static bool testRunning; ///< true if any test is currently running
cudaEvent_t beginEvent; ///< begin CUDA event
cudaEvent_t endEvent; ///< end CUDA event
std::vector<float> times; ///< collected times
const bool disabled; ///< true if this object is disabled
public:
/// Checks CUDA related error.
/// @param status return code to be checked
/// @param message message to be shown if there was an error
/// @return true if there was no error, false otherwise
static bool check(const cudaError_t & status, const char * message) {
#if defined(GPU_DWT_TESTING)
if((!testRunning) && status != cudaSuccess) {
const char * errorString = cudaGetErrorString(status);
fprintf(stderr, "CUDA ERROR: '%s': %s\n", message, errorString);
fflush(stderr);
return false;
}
#endif // GPU_DWT_TESTING
return true;
} }
/// Checks last kernel call for errors.
/// @param message description of the kernel call
/// @return true if there was no error, false otherwise
static bool checkLastKernelCall(const char * message) {
#if defined(GPU_DWT_TESTING)
return testRunning ? true : check(cudaThreadSynchronize(), message);
#else // GPU_DWT_TESTING
return true;
#endif // GPU_DWT_TESTING
}
/// Initializes DWT tester for time measurement
CudaDWTTester() : disabled(testRunning) {}
/// Gets rpefered number of iterations
int getNumIterations() {
return disabled ? 1 : 31;
}
/// Starts one test iteration.
void beginTestIteration() {
if(!disabled) {
cudaEventCreate(&beginEvent);
cudaEventCreate(&endEvent);
cudaEventRecord(beginEvent, 0);
testRunning = true;
}
}
/// Ends on etest iteration.
void endTestIteration() {
if(!disabled) {
float time;
testRunning = false;
cudaEventRecord(endEvent, 0);
cudaEventSynchronize(endEvent);
cudaEventElapsedTime(&time, beginEvent, endEvent);
cudaEventDestroy(beginEvent);
cudaEventDestroy(endEvent);
times.push_back(time);
}
}
/// Shows brief info about all iterations.
/// @param name name of processing method
/// @param sizeX width of processed image
/// @param sizeY height of processed image
void showPerformance(const char * name, const int sizeX, const int sizeY) {
if(!disabled) {
// compute mean and median
std::sort(times.begin(), times.end());
double sum = 0;
for(int i = times.size(); i--; ) {
sum += times[i];
}
const double median = (times[times.size() / 2]
+ times[(times.size() - 1) / 2]) * 0.5f;
printf(" %s: %7.3f ms (mean) %7.3f ms (median) %7.3f ms (max) "
"(%d x %d)\n", name, (sum / times.size()), median,
times[times.size() - 1], sizeX, sizeY);
}
}
};
/// Simple cudaMemcpy wrapped in performance tester.
/// @param dest destination bufer
/// @param src source buffer
/// @param sx width of copied image
/// @param sy height of copied image
template <typename T>
inline void memCopy(T * const dest, const T * const src,
const size_t sx, const size_t sy) {
cudaError_t status;
PERF_BEGIN
status = cudaMemcpy(dest, src, sx*sy*sizeof(T), cudaMemcpyDeviceToDevice);
PERF_END(" memcpy", sx, sy)
CudaDWTTester::check(status, "memcpy device > device");
} }
/// Ends on etest iteration.
void endTestIteration() {
if (!disabled) {
float time;
testRunning = false;
cudaEventRecord(endEvent, 0);
cudaEventSynchronize(endEvent);
cudaEventElapsedTime(&time, beginEvent, endEvent);
cudaEventDestroy(beginEvent);
cudaEventDestroy(endEvent);
times.push_back(time);
}
}
/// Shows brief info about all iterations.
/// @param name name of processing method
/// @param sizeX width of processed image
/// @param sizeY height of processed image
void showPerformance(const char *name, const int sizeX, const int sizeY) {
if (!disabled) {
// compute mean and median
std::sort(times.begin(), times.end());
double sum = 0;
for (int i = times.size(); i--;) {
sum += times[i];
}
const double median =
(times[times.size() / 2] + times[(times.size() - 1) / 2]) * 0.5f;
printf(" %s: %7.3f ms (mean) %7.3f ms (median) %7.3f ms (max) "
"(%d x %d)\n",
name, (sum / times.size()), median, times[times.size() - 1], sizeX,
sizeY);
}
}
};
/// Simple cudaMemcpy wrapped in performance tester.
/// @param dest destination bufer
/// @param src source buffer
/// @param sx width of copied image
/// @param sy height of copied image
template <typename T>
inline void memCopy(T *const dest, const T *const src, const size_t sx,
const size_t sy) {
cudaError_t status;
PERF_BEGIN
status = cudaMemcpy(dest, src, sx * sy * sizeof(T), cudaMemcpyDeviceToDevice);
PERF_END(" memcpy", sx, sy)
CudaDWTTester::check(status, "memcpy device > device");
}
} // end of namespace dwt_cuda } // end of namespace dwt_cuda
#endif // DWT_COMMON_CUDA_H
#endif // DWT_COMMON_CUDA_H

81
examples/dwt2d/dwt_cuda/dwt.h Executable file → Normal file
View File

@ -56,57 +56,48 @@
/// ///
#ifndef DWT_CUDA_H #ifndef DWT_CUDA_H
#define DWT_CUDA_H #define DWT_CUDA_H
namespace dwt_cuda { namespace dwt_cuda {
/// Forward 5/3 2D DWT. See common rules (above) for more details.
/// @param in Expected to be normalized into range [-128, 127].
/// Will not be preserved (will be overwritten).
/// @param out output buffer on GPU
/// @param sizeX width of input image (in pixels)
/// @param sizeY height of input image (in pixels)
/// @param levels number of recursive DWT levels
void fdwt53(int *in, int *out, int sizeX, int sizeY, int levels);
/// Forward 5/3 2D DWT. See common rules (above) for more details. /// Reverse 5/3 2D DWT. See common rules (above) for more details.
/// @param in Expected to be normalized into range [-128, 127]. /// @param in Input DWT coefficients. Format described in common rules.
/// Will not be preserved (will be overwritten). /// Will not be preserved (will be overwritten).
/// @param out output buffer on GPU /// @param out output buffer on GPU - will contain original image
/// @param sizeX width of input image (in pixels) /// in normalized range [-128, 127].
/// @param sizeY height of input image (in pixels) /// @param sizeX width of input image (in pixels)
/// @param levels number of recursive DWT levels /// @param sizeY height of input image (in pixels)
void fdwt53(int * in, int * out, int sizeX, int sizeY, int levels); /// @param levels number of recursive DWT levels
void rdwt53(int *in, int *out, int sizeX, int sizeY, int levels);
/// Forward 9/7 2D DWT. See common rules (above) for more details.
/// @param in Input DWT coefficients. Should be normalized (in range
/// [-0.5, 0.5]). Will not be preserved (will be overwritten).
/// @param out output buffer on GPU - format specified in common rules
/// @param sizeX width of input image (in pixels)
/// @param sizeY height of input image (in pixels)
/// @param levels number of recursive DWT levels
void fdwt97(float *in, float *out, int sizeX, int sizeY, int levels);
/// Reverse 5/3 2D DWT. See common rules (above) for more details. /// Reverse 9/7 2D DWT. See common rules (above) for more details.
/// @param in Input DWT coefficients. Format described in common rules. /// @param in Input DWT coefficients. Format described in common rules.
/// Will not be preserved (will be overwritten). /// Will not be preserved (will be overwritten).
/// @param out output buffer on GPU - will contain original image /// @param out output buffer on GPU - will contain original image
/// in normalized range [-128, 127]. /// in normalized range [-0.5, 0.5].
/// @param sizeX width of input image (in pixels) /// @param sizeX width of input image (in pixels)
/// @param sizeY height of input image (in pixels) /// @param sizeY height of input image (in pixels)
/// @param levels number of recursive DWT levels /// @param levels number of recursive DWT levels
void rdwt53(int * in, int * out, int sizeX, int sizeY, int levels); void rdwt97(float *in, float *out, int sizeX, int sizeY, int levels);
/// Forward 9/7 2D DWT. See common rules (above) for more details.
/// @param in Input DWT coefficients. Should be normalized (in range
/// [-0.5, 0.5]). Will not be preserved (will be overwritten).
/// @param out output buffer on GPU - format specified in common rules
/// @param sizeX width of input image (in pixels)
/// @param sizeY height of input image (in pixels)
/// @param levels number of recursive DWT levels
void fdwt97(float * in, float * out, int sizeX, int sizeY, int levels);
/// Reverse 9/7 2D DWT. See common rules (above) for more details.
/// @param in Input DWT coefficients. Format described in common rules.
/// Will not be preserved (will be overwritten).
/// @param out output buffer on GPU - will contain original image
/// in normalized range [-0.5, 0.5].
/// @param sizeX width of input image (in pixels)
/// @param sizeY height of input image (in pixels)
/// @param levels number of recursive DWT levels
void rdwt97(float * in, float * out, int sizeX, int sizeY, int levels);
} // namespace dwt_cuda } // namespace dwt_cuda
#endif // DWT_CUDA_H
#endif // DWT_CUDA_H

763
examples/dwt2d/dwt_cuda/io.h Executable file → Normal file
View File

@ -30,454 +30,411 @@
/// POSSIBILITY OF SUCH DAMAGE. /// POSSIBILITY OF SUCH DAMAGE.
/// ///
#ifndef IO_H #ifndef IO_H
#define IO_H #define IO_H
#include "common.h" #include "common.h"
namespace dwt_cuda { namespace dwt_cuda {
/// Base for all IO classes - manages mirroring.
/// Base for all IO classes - manages mirroring. class DWTIO {
class DWTIO { protected:
protected: /// Handles mirroring of image at edges in a DWT correct way.
/// Handles mirroring of image at edges in a DWT correct way. /// @param d a position in the image (will be replaced by mirrored d)
/// @param d a position in the image (will be replaced by mirrored d) /// @param sizeD size of the image along the dimension of 'd'
/// @param sizeD size of the image along the dimension of 'd' __device__ static void mirror(int &d, const int &sizeD) {
__device__ static void mirror(int & d, const int & sizeD) { // TODO: enable multiple mirroring:
// TODO: enable multiple mirroring: // if(sizeD > 1) {
// if(sizeD > 1) { // if(d < 0) {
// if(d < 0) { // const int underflow = -1 - d;
// const int underflow = -1 - d; // const int phase = (underflow / (sizeD - 1)) & 1;
// const int phase = (underflow / (sizeD - 1)) & 1; // const int remainder = underflow % (sizeD - 1);
// const int remainder = underflow % (sizeD - 1); // if(phase == 0) {
// if(phase == 0) { // d = remainder + 1;
// d = remainder + 1; // } else {
// } else { // d = sizeD - 2 - remainder;
// d = sizeD - 2 - remainder; // }
// } // } else if(d >= sizeD) {
// } else if(d >= sizeD) { // const int overflow = d - sizeD;
// const int overflow = d - sizeD; // const int phase = (overflow / (sizeD - 1)) & 1;
// const int phase = (overflow / (sizeD - 1)) & 1; // const int remainder = overflow % (sizeD - 1);
// const int remainder = overflow % (sizeD - 1); // if(phase == 0) {
// if(phase == 0) { // d = sizeD - 2 - remainder;
// d = sizeD - 2 - remainder; // } else {
// } else { // d = remainder + 1;
// d = remainder + 1; // }
// } // }
// } // } else {
// } else { // d = 0;
// d = 0; // }
// } // for test the mirror's use Feb 17
//for test the mirror's use Feb 17 if (d >= sizeD) {
if(d >= sizeD) { d = 2 * sizeD - 2 - d;
d = 2 * sizeD - 2 - d; } else if (d < 0) {
} else if(d < 0) { d = -d;
d = -d;
}
} }
}; }
};
/// Base class for pixel loader and writer - manages computing start index,
/// stride and end of image for loading column of pixels.
/// @tparam T type of image pixels
/// @tparam CHECKED true = be prepared to image boundary, false = don't care
template <typename T, bool CHECKED> class VerticalDWTPixelIO : protected DWTIO {
protected:
int end; ///< index of bottom neightbor of last pixel of column
int stride; ///< increment of pointer to get to next pixel
/// Base class for pixel loader and writer - manages computing start index, /// Initializes pixel IO - sets end index and a position of first pixel.
/// stride and end of image for loading column of pixels. /// @param sizeX width of the image
/// @tparam T type of image pixels /// @param sizeY height of the image
/// @tparam CHECKED true = be prepared to image boundary, false = don't care /// @param firstX x-coordinate of first pixel to use
template <typename T, bool CHECKED> /// @param firstY y-coordinate of first pixel to use
class VerticalDWTPixelIO : protected DWTIO { /// @return index of pixel at position [x, y] in the image
protected: __device__ int initialize(const int sizeX, const int sizeY, int firstX,
int end; ///< index of bottom neightbor of last pixel of column int firstY) {
int stride; ///< increment of pointer to get to next pixel // initialize all pointers and stride
end = CHECKED ? (sizeY * sizeX + firstX) : 0;
stride = sizeX;
return firstX + sizeX * firstY;
}
};
/// Initializes pixel IO - sets end index and a position of first pixel. /// Writes reverse transformed pixels directly into output image.
/// @param sizeX width of the image /// @tparam T type of output pixels
/// @param sizeY height of the image /// @tparam CHECKED true = be prepared to image boundary, false = don't care
/// @param firstX x-coordinate of first pixel to use template <typename T, bool CHECKED>
/// @param firstY y-coordinate of first pixel to use class VerticalDWTPixelWriter : VerticalDWTPixelIO<T, CHECKED> {
/// @return index of pixel at position [x, y] in the image private:
__device__ int initialize(const int sizeX, const int sizeY, int next; // index of the next pixel to be loaded
int firstX, int firstY) {
// initialize all pointers and stride
end = CHECKED ? (sizeY * sizeX + firstX) : 0;
stride = sizeX;
return firstX + sizeX * firstY;
}
};
public:
/// Initializes writer - sets output buffer and a position of first pixel.
/// Writes reverse transformed pixels directly into output image. /// @param sizeX width of the image
/// @tparam T type of output pixels /// @param sizeY height of the image
/// @tparam CHECKED true = be prepared to image boundary, false = don't care /// @param firstX x-coordinate of first pixel to write into
template <typename T, bool CHECKED> /// @param firstY y-coordinate of first pixel to write into
class VerticalDWTPixelWriter : VerticalDWTPixelIO<T, CHECKED> { __device__ void init(const int sizeX, const int sizeY, int firstX,
private: int firstY) {
int next; // index of the next pixel to be loaded if (firstX < sizeX) {
next = this->initialize(sizeX, sizeY, firstX, firstY);
public: } else {
/// Initializes writer - sets output buffer and a position of first pixel.
/// @param sizeX width of the image
/// @param sizeY height of the image
/// @param firstX x-coordinate of first pixel to write into
/// @param firstY y-coordinate of first pixel to write into
__device__ void init(const int sizeX, const int sizeY,
int firstX, int firstY) {
if(firstX < sizeX) {
next = this->initialize(sizeX, sizeY, firstX, firstY);
} else {
this->end = 0;
this->stride = 0;
next = 0;
}
}
/// Writes given value at next position and advances internal pointer while
/// correctly handling mirroring.
/// @param output output image to write pixel into
/// @param value value of the pixel to be written
__device__ void writeInto(T * const output, const T & value) {
if((!CHECKED) || (next != this->end)) {
output[next] = value;
next += this->stride;
}
}
};
/// Loads pixels from input image.
/// @tparam T type of image input pixels
/// @tparam CHECKED true = be prepared to image boundary, false = don't care
template <typename T, bool CHECKED>
class VerticalDWTPixelLoader
: protected VerticalDWTPixelIO<const T, CHECKED> {
private:
int last; ///< index of last loaded pixel
public:
//******************* FOR TEST **********************
__device__ int getlast(){
return last;
}
__device__ int getend(){
return this->end;
}
__device__ int getstride(){
return this->stride;
}
__device__ void setend(int a){
this->end=a;
}
//******************* FOR TEST **********************
/// Initializes loader - sets input size and a position of first pixel.
/// @param sizeX width of the image
/// @param sizeY height of the image
/// @param firstX x-coordinate of first pixel to load
/// @param firstY y-coordinate of first pixel to load
__device__ void init(const int sizeX, const int sizeY,
int firstX, int firstY) {
// correctly mirror x coordinate
this->mirror(firstX, sizeX);
// 'last' always points to already loaded pixel (subtract sizeX = stride)
last = this->initialize(sizeX, sizeY, firstX, firstY) - sizeX;
//last = (FirstX + sizeX * FirstY) - sizeX
}
/// Sets all fields to zeros, for compiler not to complain about
/// uninitialized stuff.
__device__ void clear() {
this->end = 0; this->end = 0;
this->stride = 0; this->stride = 0;
this->last = 0; next = 0;
}
}
/// Writes given value at next position and advances internal pointer while
/// correctly handling mirroring.
/// @param output output image to write pixel into
/// @param value value of the pixel to be written
__device__ void writeInto(T *const output, const T &value) {
if ((!CHECKED) || (next != this->end)) {
output[next] = value;
next += this->stride;
}
}
};
/// Loads pixels from input image.
/// @tparam T type of image input pixels
/// @tparam CHECKED true = be prepared to image boundary, false = don't care
template <typename T, bool CHECKED>
class VerticalDWTPixelLoader : protected VerticalDWTPixelIO<const T, CHECKED> {
private:
int last; ///< index of last loaded pixel
public:
//******************* FOR TEST **********************
__device__ int getlast() { return last; }
__device__ int getend() { return this->end; }
__device__ int getstride() { return this->stride; }
__device__ void setend(int a) { this->end = a; }
//******************* FOR TEST **********************
/// Initializes loader - sets input size and a position of first pixel.
/// @param sizeX width of the image
/// @param sizeY height of the image
/// @param firstX x-coordinate of first pixel to load
/// @param firstY y-coordinate of first pixel to load
__device__ void init(const int sizeX, const int sizeY, int firstX,
int firstY) {
// correctly mirror x coordinate
this->mirror(firstX, sizeX);
// 'last' always points to already loaded pixel (subtract sizeX = stride)
last = this->initialize(sizeX, sizeY, firstX, firstY) - sizeX;
// last = (FirstX + sizeX * FirstY) - sizeX
}
/// Sets all fields to zeros, for compiler not to complain about
/// uninitialized stuff.
__device__ void clear() {
this->end = 0;
this->stride = 0;
this->last = 0;
}
/// Gets another pixel and advancees internal pointer to following one.
/// @param input input image to load next pixel from
/// @return next pixel from given image
__device__ T loadFrom(const T *const input) {
last += this->stride;
if (CHECKED && (last == this->end)) {
last -= 2 * this->stride;
this->stride = -this->stride; // reverse loader's direction
}
// avoid reading from negative indices if loader is checked
// return (CHECKED && (last < 0)) ? 0 : input[last]; // TODO: use this
// checked variant later
if (last < 0) {
return 0;
} }
/// Gets another pixel and advancees internal pointer to following one. return input[last];
/// @param input input image to load next pixel from // return this->end;
/// @return next pixel from given image // return last;
__device__ T loadFrom(const T * const input) { // return this->stride;
last += this->stride; }
if(CHECKED && (last == this->end)) { };
last -= 2 * this->stride;
this->stride = -this->stride; // reverse loader's direction
}
// avoid reading from negative indices if loader is checked
// return (CHECKED && (last < 0)) ? 0 : input[last]; // TODO: use this checked variant later
if(last < 0 ) {
return 0;
}
return input[last]; /// Base for band write and loader. Manages computing strides and pointers
// return this->end; /// to first and last pixels in a linearly-stored-bands correct way.
// return last; /// @tparam T type of band coefficients
// return this->stride; /// @tparam CHECKED true = be prepared to image boundary, false = don't care
} template <typename T, bool CHECKED> class VerticalDWTBandIO : protected DWTIO {
}; protected:
/// index of bottom neighbor of last pixel of loaded column
int end;
/// increment of index to get from highpass band to the lowpass one
int strideHighToLow;
/// increment of index to get from the lowpass band to the highpass one
int strideLowToHigh;
/// Base for band write and loader. Manages computing strides and pointers /// Initializes IO - sets size of image and a position of first pixel.
/// to first and last pixels in a linearly-stored-bands correct way. /// @param imageSizeX width of the image
/// @tparam T type of band coefficients /// @param imageSizeY height of the image
/// @tparam CHECKED true = be prepared to image boundary, false = don't care /// @param firstX x-coordinate of first pixel to use
template <typename T, bool CHECKED> /// (Parity determines vertically low or high band.)
class VerticalDWTBandIO : protected DWTIO { /// @param firstY y-coordinate of first pixel to use
protected: /// (Parity determines horizontally low or high band.)
/// index of bottom neighbor of last pixel of loaded column /// @return index of first item specified by firstX and firstY
int end; __device__ int initialize(const int imageSizeX, const int imageSizeY,
int firstX, int firstY) {
// index of first pixel (topmost one) of the column with index firstX
int columnOffset = firstX / 2;
/// increment of index to get from highpass band to the lowpass one // difference between indices of two vertically neighboring pixels
int strideHighToLow; // in the same band
int verticalStride;
/// increment of index to get from the lowpass band to the highpass one // resolve index of first pixel according to horizontal parity
int strideLowToHigh; if (firstX & 1) {
// first pixel in one of right bands
/// Initializes IO - sets size of image and a position of first pixel. verticalStride = imageSizeX / 2;
/// @param imageSizeX width of the image columnOffset += divRndUp(imageSizeX, 2) * divRndUp(imageSizeY, 2);
/// @param imageSizeY height of the image strideLowToHigh = (imageSizeX * imageSizeY) / 2;
/// @param firstX x-coordinate of first pixel to use } else {
/// (Parity determines vertically low or high band.) // first pixel in one of left bands
/// @param firstY y-coordinate of first pixel to use verticalStride = imageSizeX / 2 + (imageSizeX & 1);
/// (Parity determines horizontally low or high band.) strideLowToHigh = divRndUp(imageSizeY, 2) * imageSizeX;
/// @return index of first item specified by firstX and firstY
__device__ int initialize(const int imageSizeX, const int imageSizeY,
int firstX, int firstY) {
// index of first pixel (topmost one) of the column with index firstX
int columnOffset = firstX / 2;
// difference between indices of two vertically neighboring pixels
// in the same band
int verticalStride;
// resolve index of first pixel according to horizontal parity
if(firstX & 1) {
// first pixel in one of right bands
verticalStride = imageSizeX / 2;
columnOffset += divRndUp(imageSizeX, 2) * divRndUp(imageSizeY, 2);
strideLowToHigh = (imageSizeX * imageSizeY) / 2;
} else {
// first pixel in one of left bands
verticalStride = imageSizeX / 2 + (imageSizeX & 1);
strideLowToHigh = divRndUp(imageSizeY, 2) * imageSizeX;
}
// set the other stride
strideHighToLow = verticalStride - strideLowToHigh;
// compute index of coefficient which indicates end of image
if(CHECKED) {
end = columnOffset // right column
+ (imageSizeY / 2) * verticalStride // right row
+ (imageSizeY & 1) * strideLowToHigh; // possibly in high band
} else {
end = 0;
}
//***********for test**************
// end = CHECKED;
//***********for test**************
// finally, return index of the first item
return columnOffset // right column
+ (firstY / 2) * verticalStride // right row
+ (firstY & 1) * strideLowToHigh; // possibly in high band
}
};
/// Directly loads coefficients from four consecutively stored transformed
/// bands.
/// @tparam T type of input band coefficients
/// @tparam CHECKED true = be prepared to image boundary, false = don't care
template <typename T, bool CHECKED>
class VerticalDWTBandLoader : public VerticalDWTBandIO<const T, CHECKED> {
private:
int last; ///< index of last loaded pixel
/// Checks internal index and possibly reverses direction of loader.
/// (Handles mirroring at the bottom of the image.)
/// @param input input image to load next coefficient from
/// @param stride stride to use now (one of two loader's strides)
/// @return loaded coefficient
__device__ T updateAndLoad(const T * const input, const int & stride) {
last += stride;
if(CHECKED && (last == this->end)) {
// undo last two updates of index (to get to previous mirrored item)
last -= (this->strideLowToHigh + this->strideHighToLow);
// swap and reverse strides (to move up in the loaded column now)
const int temp = this->strideLowToHigh;
this->strideLowToHigh = -this->strideHighToLow;
this->strideHighToLow = -temp;
}
if(last < 0 ) {
return 0;
}
// avoid reading from negative indices if loader is checked
// return (CHECKED && (last < 0)) ? 0 : input[last]; // TODO: use this checked variant later
return input[last];
}
public:
/// Initializes loader - sets input size and a position of first pixel.
/// @param imageSizeX width of the image
/// @param imageSizeY height of the image
/// @param firstX x-coordinate of first pixel to load
/// (Parity determines vertically low or high band.)
/// @param firstY y-coordinate of first pixel to load
/// (Parity determines horizontally low or high band.)
__device__ void init(const int imageSizeX, const int imageSizeY,
int firstX, const int firstY) {
this->mirror(firstX, imageSizeX);
last = this->initialize(imageSizeX, imageSizeY, firstX, firstY);
// adjust to point to previous item
last -= (firstY & 1) ? this->strideLowToHigh : this->strideHighToLow;
} }
/// Sets all fields to zeros, for compiler not to complain about // set the other stride
/// uninitialized stuff. strideHighToLow = verticalStride - strideLowToHigh;
__device__ void clear() {
this->end = 0; // compute index of coefficient which indicates end of image
this->strideHighToLow = 0; if (CHECKED) {
this->strideLowToHigh = 0; end = columnOffset // right column
this->last = 0; + (imageSizeY / 2) * verticalStride // right row
+ (imageSizeY & 1) * strideLowToHigh; // possibly in high band
} else {
end = 0;
} }
/// Gets another coefficient from lowpass band and advances internal index. //***********for test**************
/// Call this method first if position of first pixel passed to init // end = CHECKED;
/// was in high band. //***********for test**************
/// @param input input image to load next coefficient from
/// @return next coefficient from the lowpass band of the given image // finally, return index of the first item
__device__ T loadLowFrom(const T * const input) { return columnOffset // right column
return updateAndLoad(input, this->strideHighToLow); + (firstY / 2) * verticalStride // right row
+ (firstY & 1) * strideLowToHigh; // possibly in high band
}
};
/// Directly loads coefficients from four consecutively stored transformed
/// bands.
/// @tparam T type of input band coefficients
/// @tparam CHECKED true = be prepared to image boundary, false = don't care
template <typename T, bool CHECKED>
class VerticalDWTBandLoader : public VerticalDWTBandIO<const T, CHECKED> {
private:
int last; ///< index of last loaded pixel
/// Checks internal index and possibly reverses direction of loader.
/// (Handles mirroring at the bottom of the image.)
/// @param input input image to load next coefficient from
/// @param stride stride to use now (one of two loader's strides)
/// @return loaded coefficient
__device__ T updateAndLoad(const T *const input, const int &stride) {
last += stride;
if (CHECKED && (last == this->end)) {
// undo last two updates of index (to get to previous mirrored item)
last -= (this->strideLowToHigh + this->strideHighToLow);
// swap and reverse strides (to move up in the loaded column now)
const int temp = this->strideLowToHigh;
this->strideLowToHigh = -this->strideHighToLow;
this->strideHighToLow = -temp;
} }
if (last < 0) {
/// Gets another coefficient from the highpass band and advances index. return 0;
/// Call this method first if position of first pixel passed to init
/// was in high band.
/// @param input input image to load next coefficient from
/// @return next coefficient from the highbass band of the given image
__device__ T loadHighFrom(const T * const input) {
return updateAndLoad(input, this->strideLowToHigh);
} }
// avoid reading from negative indices if loader is checked
// return (CHECKED && (last < 0)) ? 0 : input[last]; // TODO: use this
// checked variant later
return input[last];
}
}; public:
/// Initializes loader - sets input size and a position of first pixel.
/// @param imageSizeX width of the image
/// @param imageSizeY height of the image
/// @param firstX x-coordinate of first pixel to load
/// (Parity determines vertically low or high band.)
/// @param firstY y-coordinate of first pixel to load
/// (Parity determines horizontally low or high band.)
__device__ void init(const int imageSizeX, const int imageSizeY, int firstX,
const int firstY) {
this->mirror(firstX, imageSizeX);
last = this->initialize(imageSizeX, imageSizeY, firstX, firstY);
// adjust to point to previous item
last -= (firstY & 1) ? this->strideLowToHigh : this->strideHighToLow;
}
/// Sets all fields to zeros, for compiler not to complain about
/// uninitialized stuff.
__device__ void clear() {
this->end = 0;
this->strideHighToLow = 0;
this->strideLowToHigh = 0;
this->last = 0;
}
/// Gets another coefficient from lowpass band and advances internal index.
/// Call this method first if position of first pixel passed to init
/// was in high band.
/// @param input input image to load next coefficient from
/// @return next coefficient from the lowpass band of the given image
__device__ T loadLowFrom(const T *const input) {
return updateAndLoad(input, this->strideHighToLow);
}
/// Directly saves coefficients into four transformed bands. /// Gets another coefficient from the highpass band and advances index.
/// @tparam T type of output band coefficients /// Call this method first if position of first pixel passed to init
/// @tparam CHECKED true = be prepared to image boundary, false = don't care /// was in high band.
template <typename T, bool CHECKED> /// @param input input image to load next coefficient from
class VerticalDWTBandWriter : public VerticalDWTBandIO<T, CHECKED> { /// @return next coefficient from the highbass band of the given image
private: __device__ T loadHighFrom(const T *const input) {
int next; ///< index of last loaded pixel return updateAndLoad(input, this->strideLowToHigh);
}
};
/// Checks internal index and possibly stops the writer. /// Directly saves coefficients into four transformed bands.
/// (Handles mirroring at edges of the image.) /// @tparam T type of output band coefficients
/// @param output output buffer /// @tparam CHECKED true = be prepared to image boundary, false = don't care
/// @param item item to put into the output template <typename T, bool CHECKED>
/// @param stride increment of the pointer to get to next output index class VerticalDWTBandWriter : public VerticalDWTBandIO<T, CHECKED> {
__device__ int saveAndUpdate(T * const output, const T & item, private:
const int & stride) { int next; ///< index of last loaded pixel
// if(blockIdx.x == 0 && blockIdx.y == 11 && threadIdx.x == 0){ //test, Mar 20
if((!CHECKED) || (next != this->end)) { /// Checks internal index and possibly stops the writer.
// if(next == 4) { /// (Handles mirroring at edges of the image.)
// printf(" next: %d stride: %d val: %f \n", next, stride, item ); /// @param output output buffer
// } /// @param item item to put into the output
output[next] = item; /// @param stride increment of the pointer to get to next output index
next += stride; __device__ int saveAndUpdate(T *const output, const T &item,
} const int &stride) {
// } // if(blockIdx.x == 0 && blockIdx.y == 11 && threadIdx.x == 0){
// if((!CHECKED) || (next != this->end)) { //the real one ////test, Mar 20
// output[next] = item; if ((!CHECKED) || (next != this->end)) {
// next += stride; //stride has been test // if(next == 4) {
// printf(" next: %d stride: %d val: %f \n", next, stride, item );
// } // }
return next; output[next] = item;
next += stride;
} }
public: // }
// if((!CHECKED) || (next != this->end)) { //the real one
// output[next] = item;
// next += stride; //stride has been test
// }
return next;
}
/// Initializes writer - sets output size and a position of first pixel. public:
/// @param output output image /// Initializes writer - sets output size and a position of first pixel.
/// @param imageSizeX width of the image /// @param output output image
/// @param imageSizeY height of the image /// @param imageSizeX width of the image
/// @param firstX x-coordinate of first pixel to write /// @param imageSizeY height of the image
/// (Parity determines vertically low or high band.) /// @param firstX x-coordinate of first pixel to write
/// @param firstY y-coordinate of first pixel to write /// (Parity determines vertically low or high band.)
/// (Parity determines horizontally low or high band.) /// @param firstY y-coordinate of first pixel to write
__device__ void init(const int imageSizeX, const int imageSizeY, /// (Parity determines horizontally low or high band.)
const int firstX, const int firstY) { __device__ void init(const int imageSizeX, const int imageSizeY,
if (firstX < imageSizeX) { const int firstX, const int firstY) {
next = this->initialize(imageSizeX, imageSizeY, firstX, firstY); if (firstX < imageSizeX) {
} else { next = this->initialize(imageSizeX, imageSizeY, firstX, firstY);
clear(); } else {
} clear();
} }
}
/// Sets all fields to zeros, for compiler not to complain about /// Sets all fields to zeros, for compiler not to complain about
/// uninitialized stuff. /// uninitialized stuff.
__device__ void clear() { __device__ void clear() {
this->end = 0; this->end = 0;
this->strideHighToLow = 0; this->strideHighToLow = 0;
this->strideLowToHigh = 0; this->strideLowToHigh = 0;
this->next = 0; this->next = 0;
} }
/// Writes another coefficient into the band which was specified using /// Writes another coefficient into the band which was specified using
/// init's firstX and firstY parameters and advances internal pointer. /// init's firstX and firstY parameters and advances internal pointer.
/// Call this method first if position of first pixel passed to init /// Call this method first if position of first pixel passed to init
/// was in lowpass band. /// was in lowpass band.
/// @param output output image /// @param output output image
/// @param low lowpass coefficient to save into the lowpass band /// @param low lowpass coefficient to save into the lowpass band
__device__ int writeLowInto(T * const output, const T & primary) { __device__ int writeLowInto(T *const output, const T &primary) {
return saveAndUpdate(output, primary, this->strideLowToHigh); return saveAndUpdate(output, primary, this->strideLowToHigh);
} }
/// Writes another coefficient from the other band and advances pointer. /// Writes another coefficient from the other band and advances pointer.
/// Call this method first if position of first pixel passed to init /// Call this method first if position of first pixel passed to init
/// was in highpass band. /// was in highpass band.
/// @param output output image /// @param output output image
/// @param high highpass coefficient to save into the highpass band /// @param high highpass coefficient to save into the highpass band
__device__ int writeHighInto(T * const output, const T & other) { __device__ int writeHighInto(T *const output, const T &other) {
return saveAndUpdate(output, other, this->strideHighToLow); return saveAndUpdate(output, other, this->strideHighToLow);
} }
//*******Add three functions to get private values******* //*******Add three functions to get private values*******
__device__ int getnext(){ __device__ int getnext() { return next; }
return next;
}
__device__ int getend(){ __device__ int getend() { return this->end; }
return this->end;
}
__device__ int getstrideHighToLow(){ __device__ int getstrideHighToLow() { return this->strideHighToLow; }
return this->strideHighToLow;
}
__device__ int getstrideLowToHigh(){
return this->strideLowToHigh;
}
//*******Add three functions to get private values*******
};
__device__ int getstrideLowToHigh() { return this->strideLowToHigh; }
//*******Add three functions to get private values*******
};
} // namespace dwt_cuda } // namespace dwt_cuda
#endif // IO_H
#endif // IO_H

543
examples/dwt2d/dwt_cuda/transform_buffer.h Executable file → Normal file
View File

@ -30,344 +30,309 @@
/// POSSIBILITY OF SUCH DAMAGE. /// POSSIBILITY OF SUCH DAMAGE.
/// ///
#ifndef TRANSFORM_BUFFER_H #ifndef TRANSFORM_BUFFER_H
#define TRANSFORM_BUFFER_H #define TRANSFORM_BUFFER_H
namespace dwt_cuda { namespace dwt_cuda {
/// Buffer (in shared memory of GPU) where block of input image is stored,
/// but odd and even lines are separated. (Generates less bank conflicts when
/// using lifting schema.) All operations expect SIZE_X threads.
/// Also implements basic building blocks of lifting schema.
/// @tparam SIZE_X width of the buffer excluding two boundaries (Also
/// a number of threads participating on all operations.)
/// Must be divisible by 4.
/// @tparam SIZE_Y height of buffer (total number of lines)
/// @tparam BOUNDARY_X number of extra pixels at the left and right side
/// boundary is expected to be smaller than half SIZE_X
/// Must be divisible by 2.
template <typename T, int SIZE_X, int SIZE_Y, int BOUNDARY_X>
class TransformBuffer {
public:
enum {
/// difference between pointers to two vertical neigbors
VERTICAL_STRIDE = BOUNDARY_X + (SIZE_X / 2)
};
/// Buffer (in shared memory of GPU) where block of input image is stored, private:
/// but odd and even lines are separated. (Generates less bank conflicts when enum {
/// using lifting schema.) All operations expect SIZE_X threads. /// number of shared memory banks - needed for correct padding
/// Also implements basic building blocks of lifting schema. #ifdef __CUDA_ARCH__
/// @tparam SIZE_X width of the buffer excluding two boundaries (Also SHM_BANKS = ((__CUDA_ARCH__ >= 200) ? 32 : 16),
/// a number of threads participating on all operations.) #else
/// Must be divisible by 4. SHM_BANKS = 16, // for host code only - can be anything, won't be used
/// @tparam SIZE_Y height of buffer (total number of lines) #endif
/// @tparam BOUNDARY_X number of extra pixels at the left and right side
/// boundary is expected to be smaller than half SIZE_X
/// Must be divisible by 2.
template <typename T, int SIZE_X, int SIZE_Y, int BOUNDARY_X>
class TransformBuffer {
public:
enum {
/// difference between pointers to two vertical neigbors
VERTICAL_STRIDE = BOUNDARY_X + (SIZE_X / 2)
};
private: /// size of one of two buffers (odd or even)
enum { BUFFER_SIZE = VERTICAL_STRIDE * SIZE_Y,
/// number of shared memory banks - needed for correct padding
#ifdef __CUDA_ARCH__
SHM_BANKS = ((__CUDA_ARCH__ >= 200) ? 32 : 16),
#else
SHM_BANKS = 16, // for host code only - can be anything, won't be used
#endif
/// size of one of two buffers (odd or even) /// unused space between two buffers
BUFFER_SIZE = VERTICAL_STRIDE * SIZE_Y, PADDING = SHM_BANKS - ((BUFFER_SIZE + SHM_BANKS / 2) % SHM_BANKS),
/// unused space between two buffers /// offset of the odd columns buffer from the beginning of data buffer
PADDING = SHM_BANKS - ((BUFFER_SIZE + SHM_BANKS / 2) % SHM_BANKS), ODD_OFFSET = BUFFER_SIZE + PADDING,
};
/// offset of the odd columns buffer from the beginning of data buffer /// buffer for both even and odd columns
ODD_OFFSET = BUFFER_SIZE + PADDING, T data[2 * BUFFER_SIZE + PADDING];
};
/// buffer for both even and odd columns /// Applies specified function to all central elements while also passing
T data[2 * BUFFER_SIZE + PADDING]; /// previous and next elements as parameters.
/// @param count count of central elements to apply function to
/// @param prevOffset offset of first central element
/// @param midOffset offset of first central element's predecessor
/// @param nextOffset offset of first central element's successor
/// @param function the function itself
template <typename FUNC>
__device__ void horizontalStep(const int count, const int prevOffset,
const int midOffset, const int nextOffset,
const FUNC &function) {
// number of unchecked iterations
const int STEPS = count / SIZE_X;
// items remaining after last unchecked iteration
const int finalCount = count % SIZE_X;
// offset of items processed in last (checked) iteration
const int finalOffset = count - finalCount;
/// Applies specified function to all central elements while also passing // all threads perform fixed number of iterations ...
/// previous and next elements as parameters. for (int i = 0; i < STEPS; i++) {
/// @param count count of central elements to apply function to
/// @param prevOffset offset of first central element
/// @param midOffset offset of first central element's predecessor
/// @param nextOffset offset of first central element's successor
/// @param function the function itself
template <typename FUNC>
__device__ void horizontalStep(const int count, const int prevOffset,
const int midOffset, const int nextOffset,
const FUNC & function) {
// number of unchecked iterations
const int STEPS = count / SIZE_X;
// items remaining after last unchecked iteration
const int finalCount = count % SIZE_X;
// offset of items processed in last (checked) iteration
const int finalOffset = count - finalCount;
// all threads perform fixed number of iterations ...
for(int i = 0; i < STEPS; i++) {
// for(int i = 0; i < 3; i++) { // for(int i = 0; i < 3; i++) {
const T previous = data[prevOffset + i * SIZE_X + threadIdx.x]; const T previous = data[prevOffset + i * SIZE_X + threadIdx.x];
const T next = data[nextOffset + i * SIZE_X + threadIdx.x]; const T next = data[nextOffset + i * SIZE_X + threadIdx.x];
T & center = data[midOffset + i * SIZE_X + threadIdx.x]; T &center = data[midOffset + i * SIZE_X + threadIdx.x];
// function(previous, center, (nextOffset + i*SIZE_X+threadIdx.x)); // function(previous, center, (nextOffset + i*SIZE_X+threadIdx.x));
function(previous, center, next);// the real one function(previous, center, next); // the real one
}
// ... but not all threads participate on final iteration
if(threadIdx.x < finalCount) {
const T previous = data[prevOffset + finalOffset + threadIdx.x];
const T next = data[nextOffset + finalOffset + threadIdx.x];
T & center = data[midOffset + finalOffset + threadIdx.x];
// function(previous, center, (nextOffset+finalOffset+threadIdx.x));
// kaixi
function(previous, center, next);//the real one
}
} }
public: // ... but not all threads participate on final iteration
if (threadIdx.x < finalCount) {
__device__ void getPrintData() { const T previous = data[prevOffset + finalOffset + threadIdx.x];
// const T next = data[nextOffset + finalOffset + threadIdx.x];
for(int i = 0 ; i< 2 * BUFFER_SIZE + PADDING ; i++) { T &center = data[midOffset + finalOffset + threadIdx.x];
printf(" index: %d data: %f \n ", i ,data[i]); // function(previous, center, (nextOffset+finalOffset+threadIdx.x));
} // kaixi
function(previous, center, next); // the real one
}
/// Gets offset of the column with given index. Central columns have
/// indices from 0 to NUM_LINES - 1, left boundary columns have negative
/// indices and right boundary columns indices start with NUM_LINES.
/// @param columnIndex index of column to get pointer to
/// @return offset of the first item of column with specified index
__device__ int getColumnOffset(int columnIndex) {
columnIndex += BOUNDARY_X; // skip boundary
return columnIndex / 2 // select right column
+ (columnIndex & 1) * ODD_OFFSET; // select odd or even buffer
} }
}
public:
/// Provides access to data of the transform buffer. __device__ void getPrintData() {
/// @param index index of the item to work with //
/// @return reference to item at given index for (int i = 0; i < 2 * BUFFER_SIZE + PADDING; i++) {
__device__ T & operator[] (const int index) { printf(" index: %d data: %f \n ", i, data[i]);
return data[index];
} }
}
/// Gets offset of the column with given index. Central columns have
/// indices from 0 to NUM_LINES - 1, left boundary columns have negative
/// indices and right boundary columns indices start with NUM_LINES.
/// @param columnIndex index of column to get pointer to
/// @return offset of the first item of column with specified index
__device__ int getColumnOffset(int columnIndex) {
columnIndex += BOUNDARY_X; // skip boundary
return columnIndex / 2 // select right column
+ (columnIndex & 1) * ODD_OFFSET; // select odd or even buffer
}
/// Applies specified function to all horizontally even elements in /// Provides access to data of the transform buffer.
/// specified lines. (Including even elements in boundaries except /// @param index index of the item to work with
/// first even element in first left boundary.) SIZE_X threads participate /// @return reference to item at given index
/// and synchronization is needed before result can be used. __device__ T &operator[](const int index) { return data[index]; }
/// @param firstLine index of first line
/// @param numLines count of lines
/// @param func function to be applied on all even elements
/// parameters: previous (odd) element, the even
/// element itself and finally next (odd) element
template <typename FUNC>
__device__ void forEachHorizontalEven(const int firstLine,
const int numLines,
const FUNC & func) {
// number of even elemens to apply function to
const int count = numLines * VERTICAL_STRIDE - 1;
// offset of first even element
const int centerOffset = firstLine * VERTICAL_STRIDE + 1;
// offset of odd predecessor of first even element
const int prevOffset = firstLine * VERTICAL_STRIDE + ODD_OFFSET;
// offset of odd successor of first even element
const int nextOffset = prevOffset + 1;
// if(threadIdx.x == 0) { /// Applies specified function to all horizontally even elements in
/// specified lines. (Including even elements in boundaries except
/// first even element in first left boundary.) SIZE_X threads participate
/// and synchronization is needed before result can be used.
/// @param firstLine index of first line
/// @param numLines count of lines
/// @param func function to be applied on all even elements
/// parameters: previous (odd) element, the even
/// element itself and finally next (odd) element
template <typename FUNC>
__device__ void forEachHorizontalEven(const int firstLine, const int numLines,
const FUNC &func) {
// number of even elemens to apply function to
const int count = numLines * VERTICAL_STRIDE - 1;
// offset of first even element
const int centerOffset = firstLine * VERTICAL_STRIDE + 1;
// offset of odd predecessor of first even element
const int prevOffset = firstLine * VERTICAL_STRIDE + ODD_OFFSET;
// offset of odd successor of first even element
const int nextOffset = prevOffset + 1;
// printf("forEachHorizontalEven count %d, centerOffset %d prevOffset %d nextOffset %d \n", count, centerOffset, prevOffset, nextOffset); // if(threadIdx.x == 0) {
// }
// call generic horizontal step function // printf("forEachHorizontalEven count %d, centerOffset %d prevOffset %d
horizontalStep(count, prevOffset, centerOffset, nextOffset, func); // nextOffset %d \n", count, centerOffset, prevOffset, nextOffset);
} // }
// call generic horizontal step function
horizontalStep(count, prevOffset, centerOffset, nextOffset, func);
}
/// Applies given function to all horizontally odd elements in specified /// Applies given function to all horizontally odd elements in specified
/// lines. (Including odd elements in boundaries except last odd element /// lines. (Including odd elements in boundaries except last odd element
/// in last right boundary.) SIZE_X threads participate and synchronization /// in last right boundary.) SIZE_X threads participate and synchronization
/// is needed before result can be used. /// is needed before result can be used.
/// @param firstLine index of first line /// @param firstLine index of first line
/// @param numLines count of lines /// @param numLines count of lines
/// @param func function to be applied on all odd elements /// @param func function to be applied on all odd elements
/// parameters: previous (even) element, the odd /// parameters: previous (even) element, the odd
/// element itself and finally next (even) element /// element itself and finally next (even) element
template <typename FUNC> template <typename FUNC>
__device__ void forEachHorizontalOdd(const int firstLine, __device__ void forEachHorizontalOdd(const int firstLine, const int numLines,
const int numLines, const FUNC &func) {
const FUNC & func) { // numbet of odd elements to apply function to
// numbet of odd elements to apply function to const int count = numLines * VERTICAL_STRIDE - 1;
const int count = numLines * VERTICAL_STRIDE - 1; // offset of even predecessor of first odd element
// offset of even predecessor of first odd element const int prevOffset = firstLine * VERTICAL_STRIDE;
const int prevOffset = firstLine * VERTICAL_STRIDE; // offset of first odd element
// offset of first odd element const int centerOffset = prevOffset + ODD_OFFSET;
const int centerOffset = prevOffset + ODD_OFFSET; // offset of even successor of first odd element
// offset of even successor of first odd element const int nextOffset = prevOffset + 1;
const int nextOffset = prevOffset + 1;
// if(threadIdx.x == 0) { // if(threadIdx.x == 0) {
// printf("forEachHorizontalOdd count %d, centerOffset %d prevOffset %d nextOffset %d \n", count, centerOffset, prevOffset, nextOffset); // printf("forEachHorizontalOdd count %d, centerOffset %d prevOffset %d
// } // nextOffset %d \n", count, centerOffset, prevOffset, nextOffset);
// }
// call generic horizontal step function
horizontalStep(count, prevOffset, centerOffset, nextOffset, func);
}
// call generic horizontal step function /// Applies specified function to all even elements (except element #0)
horizontalStep(count, prevOffset, centerOffset, nextOffset, func); /// of given column. Each thread takes care of one column, so there's
} /// no need for synchronization.
/// @param columnOffset offset of thread's column
/// @param f function to be applied on all even elements
/// Applies specified function to all even elements (except element #0) /// parameters: previous (odd) element, the even
/// of given column. Each thread takes care of one column, so there's /// element itself and finally next (odd) element
/// no need for synchronization. template <typename F>
/// @param columnOffset offset of thread's column __device__ void forEachVerticalEven(const int columnOffset, const F &f) {
/// @param f function to be applied on all even elements if (SIZE_Y > 3) { // makes no sense otherwise
/// parameters: previous (odd) element, the even const int steps = SIZE_Y / 2 - 1;
/// element itself and finally next (odd) element for (int i = 0; i < steps; i++) {
template <typename F> const int row = 2 + i * 2;
__device__ void forEachVerticalEven(const int columnOffset, const F & f) {
if(SIZE_Y > 3) { // makes no sense otherwise
const int steps = SIZE_Y / 2 - 1;
for(int i = 0; i < steps; i++) {
const int row = 2 + i * 2;
const T prev = data[columnOffset + (row - 1) * VERTICAL_STRIDE];
const T next = data[columnOffset + (row + 1) * VERTICAL_STRIDE];
f(prev, data[columnOffset + row * VERTICAL_STRIDE] , next);
//--------------- FOR TEST -----------------
/* __syncthreads();
if ((blockIdx.x * blockDim.x + threadIdx.x) == 0){
diffOut[2500]++;
diffOut[diffOut[2500]] = 2;//data[columnOffset + row * VERTICAL_STRIDE];
}
__syncthreads();
*/ //--------------- FOR TEST -----------------
}
}
}
/// Applies specified function to all odd elements of given column.
/// Each thread takes care of one column, so there's no need for
/// synchronization.
/// @param columnOffset offset of thread's column
/// @param f function to be applied on all odd elements
/// parameters: previous (even) element, the odd
/// element itself and finally next (even) element
template <typename F>
__device__ void forEachVerticalOdd(const int columnOffset, const F & f) {
const int steps = (SIZE_Y - 1) / 2;
for(int i = 0; i < steps; i++) {
const int row = i * 2 + 1;
const T prev = data[columnOffset + (row - 1) * VERTICAL_STRIDE]; const T prev = data[columnOffset + (row - 1) * VERTICAL_STRIDE];
const T next = data[columnOffset + (row + 1) * VERTICAL_STRIDE]; const T next = data[columnOffset + (row + 1) * VERTICAL_STRIDE];
f(prev, data[columnOffset + row * VERTICAL_STRIDE], next);
f(prev, data[columnOffset + row * VERTICAL_STRIDE], next); //--------------- FOR TEST -----------------
/* __syncthreads();
if ((blockIdx.x * blockDim.x + threadIdx.x) == 0){
//--------------- FOR TEST ----------------- diffOut[2500]++;
/* __syncthreads(); diffOut[diffOut[2500]] = 2;//data[columnOffset +
if ((blockIdx.x * blockDim.x + threadIdx.x) == 0){ row * VERTICAL_STRIDE];
diffOut[2500]++; }
diffOut[diffOut[2500]] = 1; //data[columnOffset + row * VERTICAL_STRIDE]; __syncthreads();
} */ //--------------- FOR TEST -----------------
__syncthreads();
*/ //--------------- FOR TEST -----------------
} }
} }
}
/// Applies specified function to all odd elements of given column.
/// Each thread takes care of one column, so there's no need for
/// synchronization.
/// @param columnOffset offset of thread's column
/// @param f function to be applied on all odd elements
/// parameters: previous (even) element, the odd
/// element itself and finally next (even) element
template <typename F>
__device__ void forEachVerticalOdd(const int columnOffset, const F &f) {
const int steps = (SIZE_Y - 1) / 2;
for (int i = 0; i < steps; i++) {
const int row = i * 2 + 1;
const T prev = data[columnOffset + (row - 1) * VERTICAL_STRIDE];
const T next = data[columnOffset + (row + 1) * VERTICAL_STRIDE];
f(prev, data[columnOffset + row * VERTICAL_STRIDE], next);
/// Scales elements at specified lines. //--------------- FOR TEST -----------------
/// @param evenScale scaling factor for horizontally even elements /* __syncthreads();
/// @param oddScale scaling factor for horizontally odd elements if ((blockIdx.x * blockDim.x + threadIdx.x) == 0){
/// @param numLines number of lines, whose elements should be scaled diffOut[2500]++;
/// @param firstLine index of first line to scale elements in diffOut[diffOut[2500]] = 1; //data[columnOffset +
__device__ void scaleHorizontal(const T evenScale, const T oddScale, row * VERTICAL_STRIDE];
const int firstLine, const int numLines) { }
const int offset = firstLine * VERTICAL_STRIDE;
const int count = numLines * VERTICAL_STRIDE;
const int steps = count / SIZE_X;
const int finalCount = count % SIZE_X;
const int finalOffset = count - finalCount;
// printf("scaleHorizontal sizeX: %d offset %d, count, %d, steps, %d, finalCount %d, finalOffset %d \n", SIZE_X, offset, count, steps, finalCount, finalOffset); __syncthreads();
*/ //--------------- FOR TEST -----------------
}
}
// run iterations, whete all threads participate /// Scales elements at specified lines.
for(int i = 0; i < steps; i++) { /// @param evenScale scaling factor for horizontally even elements
data[threadIdx.x + i * SIZE_X + offset] *= evenScale; /// @param oddScale scaling factor for horizontally odd elements
// if(threadIdx.x + i * SIZE_X + offset == 531) { /// @param numLines number of lines, whose elements should be scaled
// printf("threadidx 531: %d \n", threadIdx.x); /// @param firstLine index of first line to scale elements in
// } __device__ void scaleHorizontal(const T evenScale, const T oddScale,
// if(threadIdx.x + i * SIZE_X + offset + ODD_OFFSET == 531) { const int firstLine, const int numLines) {
// printf("threadidx 531: %d \n", threadIdx.x); const int offset = firstLine * VERTICAL_STRIDE;
// } const int count = numLines * VERTICAL_STRIDE;
data[threadIdx.x + i * SIZE_X + offset + ODD_OFFSET] *= oddScale; const int steps = count / SIZE_X;
} const int finalCount = count % SIZE_X;
const int finalOffset = count - finalCount;
// some threads also finish remaining unscaled items // printf("scaleHorizontal sizeX: %d offset %d, count, %d, steps, %d,
if(threadIdx.x < finalCount) { // finalCount %d, finalOffset %d \n", SIZE_X, offset, count, steps,
data[threadIdx.x + finalOffset + offset] *= evenScale; // finalCount, finalOffset);
// if(threadIdx.x + finalOffset + offset == 531) {
// printf("threadidx 531: %d \n", threadIdx.x);
// }
// if(threadIdx.x + finalOffset + offset + ODD_OFFSET == 531) {
// printf("threadidx 531: %d \n", threadIdx.x);
// }
data[threadIdx.x + finalOffset + offset + ODD_OFFSET] *= oddScale;
}
// run iterations, whete all threads participate
for (int i = 0; i < steps; i++) {
data[threadIdx.x + i * SIZE_X + offset] *= evenScale;
// if(threadIdx.x + i * SIZE_X + offset == 531) {
// printf("threadidx 531: %d \n", threadIdx.x);
// }
// if(threadIdx.x + i * SIZE_X + offset + ODD_OFFSET == 531) {
// printf("threadidx 531: %d \n", threadIdx.x);
// }
data[threadIdx.x + i * SIZE_X + offset + ODD_OFFSET] *= oddScale;
} }
// some threads also finish remaining unscaled items
if (threadIdx.x < finalCount) {
data[threadIdx.x + finalOffset + offset] *= evenScale;
// if(threadIdx.x + finalOffset + offset == 531) {
// printf("threadidx 531: %d \n", threadIdx.x);
// }
// if(threadIdx.x + finalOffset + offset + ODD_OFFSET == 531) {
// printf("threadidx 531: %d \n", threadIdx.x);
// }
data[threadIdx.x + finalOffset + offset + ODD_OFFSET] *= oddScale;
}
}
/// Scales elements in specified column. /// Scales elements in specified column.
/// @param evenScale scaling factor for vertically even elements /// @param evenScale scaling factor for vertically even elements
/// @param oddScale scaling factor for vertically odd elements /// @param oddScale scaling factor for vertically odd elements
/// @param columnOffset offset of the column to work with /// @param columnOffset offset of the column to work with
/// @param numLines number of lines, whose elements should be scaled /// @param numLines number of lines, whose elements should be scaled
/// @param firstLine index of first line to scale elements in /// @param firstLine index of first line to scale elements in
__device__ void scaleVertical(const T evenScale, const T oddScale, __device__ void scaleVertical(const T evenScale, const T oddScale,
const int columnOffset, const int numLines, const int columnOffset, const int numLines,
const int firstLine) { const int firstLine) {
for(int i = firstLine; i < (numLines + firstLine); i++) { for (int i = firstLine; i < (numLines + firstLine); i++) {
if(i & 1) { if (i & 1) {
data[columnOffset + i * VERTICAL_STRIDE] *= oddScale; data[columnOffset + i * VERTICAL_STRIDE] *= oddScale;
} else { } else {
data[columnOffset + i * VERTICAL_STRIDE] *= evenScale; data[columnOffset + i * VERTICAL_STRIDE] *= evenScale;
}
} }
} }
}
//****************For Test(Feb23), test inter parameters*************
__device__ int getVERTICAL_STRIDE() { return VERTICAL_STRIDE; }
__device__ int getSHM_BANKS() { return SHM_BANKS; }
__device__ int getBuffersize() { return BUFFER_SIZE; }
__device__ int getPADDING() { return PADDING; }
__device__ int getODD_OFFSET() { return ODD_OFFSET; }
//****************For Test(Feb23), test inter parameters************* //****************For Test(Feb23), test inter parameters*************
__device__ int getVERTICAL_STRIDE(){
return VERTICAL_STRIDE;
}
__device__ int getSHM_BANKS(){
return SHM_BANKS;
}
__device__ int getBuffersize(){
return BUFFER_SIZE;
}
__device__ int getPADDING(){
return PADDING;
}
__device__ int getODD_OFFSET(){
return ODD_OFFSET;
}
//****************For Test(Feb23), test inter parameters*************
}; // end of class TransformBuffer
}; // end of class TransformBuffer
} // namespace dwt_cuda } // namespace dwt_cuda
#endif // TRANSFORM_BUFFER_H
#endif // TRANSFORM_BUFFER_H

1
examples/dwt2d/run_cpu.sh
View File

@ -5,4 +5,3 @@
./dwt2d 4.bmp -d 4x4 -r -5 -l 3 ./dwt2d 4.bmp -d 4x4 -r -5 -l 3
# ./dwt2d 4.bmp -d 4x4 -r -9 -l 3 # ./dwt2d 4.bmp -d 4x4 -r -9 -l 3
# ./dwt2d 8.bmp -d 8x8 -f -9 -l 3 # ./dwt2d 8.bmp -d 8x8 -f -9 -l 3

9
examples/dwt2d/test_compile_nvcc.sh
View File

@ -7,12 +7,3 @@
/usr/local/cuda/bin/nvcc -arch sm_50 -I. -I/include -O2 --compiler-options -fno-strict-aliasing -c dwt_cuda/rdwt97.cu -o dwt_cuda/rdwt97.cu.o /usr/local/cuda/bin/nvcc -arch sm_50 -I. -I/include -O2 --compiler-options -fno-strict-aliasing -c dwt_cuda/rdwt97.cu -o dwt_cuda/rdwt97.cu.o
/usr/local/cuda/bin/nvcc -arch sm_50 -I. -I/include -O2 --compiler-options -fno-strict-aliasing -c dwt_cuda/rdwt53.cu -o dwt_cuda/rdwt53.cu.o /usr/local/cuda/bin/nvcc -arch sm_50 -I. -I/include -O2 --compiler-options -fno-strict-aliasing -c dwt_cuda/rdwt53.cu -o dwt_cuda/rdwt53.cu.o
g++ -fPIC -o nvcc_dwt2d main.cu.o dwt.cu.o components.cu.o dwt_cuda/fdwt53.cu.o dwt_cuda/fdwt97.cu.o dwt_cuda/common.cu.o dwt_cuda/rdwt97.cu.o dwt_cuda/rdwt53.cu.o -L/usr/local/cuda/lib64 -lcudart g++ -fPIC -o nvcc_dwt2d main.cu.o dwt.cu.o components.cu.o dwt_cuda/fdwt53.cu.o dwt_cuda/fdwt97.cu.o dwt_cuda/common.cu.o dwt_cuda/rdwt97.cu.o dwt_cuda/rdwt53.cu.o -L/usr/local/cuda/lib64 -lcudart

30
examples/microbench/cudamemcpy_test.cc
View File

@ -1,38 +1,36 @@
#include <stdio.h> #include <stdio.h>
__global__ __global__ void saxpy(int n, float a, float *x, float *y) {
void saxpy(int n, float a, float *x, float *y) int i = blockIdx.x * blockDim.x + threadIdx.x;
{ if (i < n)
int i = blockIdx.x*blockDim.x + threadIdx.x; y[i] = a * x[i] + y[i];
if (i < n) y[i] = a*x[i] + y[i];
} }
int main(void) int main(void) {
{ int N = 1 << 20;
int N = 1<<20;
float *x, *y, *d_x, *d_y; float *x, *y, *d_x, *d_y;
x = (float*)malloc(N*sizeof(float)); x = (float *)malloc(N * sizeof(float));
y = (float*)malloc(N*sizeof(float)); y = (float *)malloc(N * sizeof(float));
cudaMalloc(&d_x, N*sizeof(float)); cudaMalloc(&d_x, N * sizeof(float));
cudaMalloc(&d_y, N*sizeof(float)); cudaMalloc(&d_y, N * sizeof(float));
for (int i = 0; i < N; i++) { for (int i = 0; i < N; i++) {
x[i] = 1.0f; x[i] = 1.0f;
y[i] = 2.0f; y[i] = 2.0f;
} }
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_x, x, N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_y, y, N * sizeof(float), cudaMemcpyHostToDevice);
// Perform SAXPY on 1M elements // Perform SAXPY on 1M elements
// saxpy<<<(N+255)/256, 256>>>(N, 2.0f, d_x, d_y); // saxpy<<<(N+255)/256, 256>>>(N, 2.0f, d_x, d_y);
cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost); cudaMemcpy(y, d_y, N * sizeof(float), cudaMemcpyDeviceToHost);
float maxError = 0.0f; float maxError = 0.0f;
for (int i = 0; i < N; i++) for (int i = 0; i < N; i++)
maxError = max(maxError, abs(y[i]-4.0f)); maxError = max(maxError, abs(y[i] - 4.0f));
printf("Max error: %f\n", maxError); printf("Max error: %f\n", maxError);
cudaFree(d_x); cudaFree(d_x);

29
examples/microbench/dummy_kernel.cc
View File

@ -1,38 +1,35 @@
#include <stdio.h> #include <stdio.h>
__global__ __global__ void saxpy(void) {
void saxpy(void) int i = blockIdx.x * blockDim.x + threadIdx.x;
{
int i = blockIdx.x*blockDim.x + threadIdx.x;
printf("block_id:%d thread_id:%d \n", i) printf("block_id:%d thread_id:%d \n", i)
} }
int main(void) int main(void) {
{ int N = 1 << 20;
int N = 1<<20;
float *x, *y, *d_x, *d_y; float *x, *y, *d_x, *d_y;
x = (float*)malloc(N*sizeof(float)); x = (float *)malloc(N * sizeof(float));
y = (float*)malloc(N*sizeof(float)); y = (float *)malloc(N * sizeof(float));
cudaMalloc(&d_x, N*sizeof(float)); cudaMalloc(&d_x, N * sizeof(float));
cudaMalloc(&d_y, N*sizeof(float)); cudaMalloc(&d_y, N * sizeof(float));
for (int i = 0; i < N; i++) { for (int i = 0; i < N; i++) {
x[i] = 1.0f; x[i] = 1.0f;
y[i] = 2.0f; y[i] = 2.0f;
} }
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_x, x, N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_y, y, N * sizeof(float), cudaMemcpyHostToDevice);
// Perform SAXPY on 1M elements // Perform SAXPY on 1M elements
saxpy<<<(1,1)>>>; saxpy<<<(1, 1)>>>;
cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost); cudaMemcpy(y, d_y, N * sizeof(float), cudaMemcpyDeviceToHost);
float maxError = 0.0f; float maxError = 0.0f;
for (int i = 0; i < N; i++) for (int i = 0; i < N; i++)
maxError = max(maxError, abs(y[i]-4.0f)); maxError = max(maxError, abs(y[i] - 4.0f));
printf("Max error: %f\n", maxError); printf("Max error: %f\n", maxError);
cudaFree(d_x); cudaFree(d_x);

29
examples/microbench/kerne_arg.cc
View File

@ -1,37 +1,32 @@
#include <stdio.h> #include <stdio.h>
__global__ __global__ void saxpy(int N) { printf("hello!: %d\n", N); }
void saxpy(int N)
{
printf("hello!: %d\n", N);
}
int main(void) int main(void) {
{ int N = 1 << 20;
int N = 1<<20;
float *x, *y, *d_x, *d_y; float *x, *y, *d_x, *d_y;
x = (float*)malloc(N*sizeof(float)); x = (float *)malloc(N * sizeof(float));
y = (float*)malloc(N*sizeof(float)); y = (float *)malloc(N * sizeof(float));
cudaMalloc(&d_x, N*sizeof(float)); cudaMalloc(&d_x, N * sizeof(float));
cudaMalloc(&d_y, N*sizeof(float)); cudaMalloc(&d_y, N * sizeof(float));
for (int i = 0; i < N; i++) { for (int i = 0; i < N; i++) {
x[i] = 1.0f; x[i] = 1.0f;
y[i] = 2.0f; y[i] = 2.0f;
} }
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_x, x, N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_y, y, N * sizeof(float), cudaMemcpyHostToDevice);
// Perform SAXPY on 1M elements // Perform SAXPY on 1M elements
saxpy<<<(1,1)>>>(N); saxpy<<<(1, 1)>>>(N);
cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost); cudaMemcpy(y, d_y, N * sizeof(float), cudaMemcpyDeviceToHost);
float maxError = 0.0f; float maxError = 0.0f;
for (int i = 0; i < N; i++) for (int i = 0; i < N; i++)
maxError = max(maxError, abs(y[i]-4.0f)); maxError = max(maxError, abs(y[i] - 4.0f));
printf("Max error: %f\n", maxError); printf("Max error: %f\n", maxError);
cudaFree(d_x); cudaFree(d_x);

29
examples/microbench/one_thread_kernel.cc
View File

@ -1,37 +1,32 @@
#include <stdio.h> #include <stdio.h>
__global__ __global__ void saxpy(void) { printf("hello!\n"); }
void saxpy(void)
{
printf("hello!\n");
}
int main(void) int main(void) {
{ int N = 1 << 20;
int N = 1<<20;
float *x, *y, *d_x, *d_y; float *x, *y, *d_x, *d_y;
x = (float*)malloc(N*sizeof(float)); x = (float *)malloc(N * sizeof(float));
y = (float*)malloc(N*sizeof(float)); y = (float *)malloc(N * sizeof(float));
cudaMalloc(&d_x, N*sizeof(float)); cudaMalloc(&d_x, N * sizeof(float));
cudaMalloc(&d_y, N*sizeof(float)); cudaMalloc(&d_y, N * sizeof(float));
for (int i = 0; i < N; i++) { for (int i = 0; i < N; i++) {
x[i] = 1.0f; x[i] = 1.0f;
y[i] = 2.0f; y[i] = 2.0f;
} }
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_x, x, N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice); cudaMemcpy(d_y, y, N * sizeof(float), cudaMemcpyHostToDevice);
// Perform SAXPY on 1M elements // Perform SAXPY on 1M elements
saxpy<<<(1,1)>>>; saxpy<<<(1, 1)>>>;
cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost); cudaMemcpy(y, d_y, N * sizeof(float), cudaMemcpyDeviceToHost);
float maxError = 0.0f; float maxError = 0.0f;
for (int i = 0; i < N; i++) for (int i = 0; i < N; i++)
maxError = max(maxError, abs(y[i]-4.0f)); maxError = max(maxError, abs(y[i] - 4.0f));
printf("Max error: %f\n", maxError); printf("Max error: %f\n", maxError);
cudaFree(d_x); cudaFree(d_x);

2
runtime/lib/cudaRuntimeImpl.cpp
View File

@ -43,7 +43,7 @@ cudaError_t cudaMallocHost(void **devPtr, size_t size) {
*devPtr = malloc(size); *devPtr = malloc(size);
if (devPtr == NULL) if (devPtr == NULL)
return cudaErrorMemoryAllocation; return cudaErrorMemoryAllocation;
return cudaSuccess; return cudaSuccess;
} }
cudaError_t cudaMemset(void *devPtr, int value, size_t count) { cudaError_t cudaMemset(void *devPtr, int value, size_t count) {
memset(devPtr, value, count); memset(devPtr, value, count);