cuda npp库旋转图片

 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <assert.h>
 #include <math.h>
 #include <windows.h>
 #include <FreeImage.h>
 #include <cuda_runtime.h>
 #include <npp.h>

 #pragma comment(lib, "FreeImage.lib")
 #pragma comment(lib, "cudart.lib")
 #pragma comment(lib, "nppi.lib")

 FIBITMAP* LoadImg( const char* szFile )
 {
     FREE_IMAGE_FORMAT nFif;

     if ( szFile == NULL || *szFile ==  )
     {
         return NULL;
     }

     if ( ( nFif = FreeImage_GetFileType( szFile,  ) ) == FIF_UNKNOWN )
     {
         if ( ( nFif = FreeImage_GetFIFFromFilename( szFile ) ) == FIF_UNKNOWN )
         {
             return NULL;
         }
     }

     if ( !FreeImage_FIFSupportsReading( nFif ) )
     {
         return NULL;
     }

     return FreeImage_Load( nFif, szFile );
 }

 int main( int argc, char* argv[] )
 {
     cudaError_t cuRet;
     NppStatus nppRet;
     BOOL fiRet;
     FIBITMAP* pSrcBmp;
     FIBITMAP* pDstBmp;
     unsigned char* pSrcData;
     unsigned char* pDstData;
     Npp8u* pSrcDataCUDA;
     Npp8u* pDstDataCUDA;
     NppiSize oSrcSize;
     NppiSize oDstSize;
     NppiRect oSrcROI;
     NppiRect oDstROI;
     int nImgBpp;
     int nSrcPitch;
     int nDstPitch;
     int nSrcPitchCUDA;
     int nDstPitchCUDA;
     double aBoundingBox[][];
     double nAngle;

     if ( argc <  )
     {
         return -;
     }

     /* 载入文件 */
     pSrcBmp = LoadImg( argv[] );
     assert( pSrcBmp != NULL );

     nImgBpp = ( FreeImage_GetBPP( pSrcBmp ) >>  );
     pSrcData = FreeImage_GetBits( pSrcBmp );

     oSrcSize.width = ( int )FreeImage_GetWidth( pSrcBmp );
     oSrcSize.height = ( int )FreeImage_GetHeight( pSrcBmp );
     nSrcPitch = ( int )FreeImage_GetPitch( pSrcBmp );

     oSrcROI.x = oSrcROI.y = ;
     oSrcROI.width = oSrcSize.width;
     oSrcROI.height = oSrcSize.height;

     nAngle = atof( argv[] );

     /* 设置显卡,构建上下文 */
     cuRet = cudaSetDevice(  );
     assert( cuRet == cudaSuccess );

     /* 分配显存 */
     switch ( nImgBpp )
     {
         case :
             pSrcDataCUDA = nppiMalloc_8u_C1( oSrcSize.width, oSrcSize.height, &nSrcPitchCUDA );
             break;
         case :
             pSrcDataCUDA = nppiMalloc_8u_C3( oSrcSize.width, oSrcSize.height, &nSrcPitchCUDA );
             break;
         case :
             pSrcDataCUDA = nppiMalloc_8u_C4( oSrcSize.width, oSrcSize.height, &nSrcPitchCUDA );
             break;
         default:
             assert(  );
             break;
     }
     assert( pSrcDataCUDA != NULL );

     /* 将原图传入显存 */
     cudaMemcpy2D( pSrcDataCUDA, nSrcPitchCUDA, pSrcData, nSrcPitch, oSrcSize.width * nImgBpp, oSrcSize.height, cudaMemcpyHostToDevice );

     /* 计算旋转后长宽 */
     nppiGetRotateBound( oSrcROI, aBoundingBox, nAngle, ,  );
     oDstSize.width = ( int )ceil( fabs( aBoundingBox[][] - aBoundingBox[][] ) );
     oDstSize.height = ( int )ceil( fabs( aBoundingBox[][] - aBoundingBox[][] ) );

     /* 建目标图 */
     pDstBmp = FreeImage_Allocate( oDstSize.width, oDstSize.height, nImgBpp <<  );
     assert( pDstBmp != NULL );

     pDstData = FreeImage_GetBits( pDstBmp );

     nDstPitch = ( int )FreeImage_GetPitch( pDstBmp );
     oDstROI.x = oDstROI.y = ;
     oDstROI.width = oDstSize.width;
     oDstROI.height = oDstSize.height;

     /* 分配显存 */
     switch ( nImgBpp )
     {
         case :
             pDstDataCUDA = nppiMalloc_8u_C1( oDstSize.width, oDstSize.height, &nDstPitchCUDA );
             break;
         case :
             pDstDataCUDA = nppiMalloc_8u_C3( oDstSize.width, oDstSize.height, &nDstPitchCUDA );
             break;
         case :
             pDstDataCUDA = nppiMalloc_8u_C4( oDstSize.width, oDstSize.height, &nDstPitchCUDA );
             break;
     }
     assert( pDstDataCUDA != NULL );
     cudaMemset2D( pDstDataCUDA, nDstPitchCUDA, , oDstSize.width * nImgBpp, oDstSize.height );

     /* 处理 */
     switch ( nImgBpp )
     {
         case :
             nppRet = nppiRotate_8u_C1R( pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
                                         pDstDataCUDA, nDstPitchCUDA, oDstROI,
                                         nAngle, - aBoundingBox[][], - aBoundingBox[][], NPPI_INTER_CUBIC );
             break;
         case :
             nppRet = nppiRotate_8u_C3R( pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
                                         pDstDataCUDA, nDstPitchCUDA, oDstROI,
                                         nAngle, - aBoundingBox[][], - aBoundingBox[][], NPPI_INTER_CUBIC );
             break;
         case :
             nppRet = nppiRotate_8u_C4R( pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
                                         pDstDataCUDA, nDstPitchCUDA, oDstROI,
                                         nAngle, - aBoundingBox[][], - aBoundingBox[][], NPPI_INTER_CUBIC );
             break;
     }
     assert( nppRet == NPP_NO_ERROR );

     cudaMemcpy2D( pDstData, nDstPitch, pDstDataCUDA, nDstPitchCUDA, oDstSize.width * nImgBpp, oDstSize.height, cudaMemcpyDeviceToHost );

     fiRet = FreeImage_Save( FIF_BMP, pDstBmp, "ret.bmp" );
     assert( fiRet );

     nppiFree( pSrcDataCUDA );
     nppiFree( pDstDataCUDA );

     cudaDeviceReset();

     FreeImage_Unload( pSrcBmp );
     FreeImage_Unload( pDstBmp );

     return ;
 }

下面是所有通道的，扩展了一点。以后自己用

 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <assert.h>
 #include <math.h>
 #include <windows.h>

 #include <FreeImage.h>

 #include <cuda_runtime.h>
 #include <npp.h>

 int main( int argc, char* argv[] )
 {
     cudaError_t cuRet;
     NppStatus nppRet;
     BOOL fiRet;
     FIBITMAP* pSrcBmp;
     FIBITMAP* pDstBmp;
     unsigned char* pSrcData;
     unsigned char* pDstData;
     Npp8u* pSrcDataCUDA;
     Npp8u* pDstDataCUDA;
     NppiSize oSrcSize;
     NppiSize oDstSize;
     NppiRect oSrcROI;
     NppiRect oDstROI;
     int nSrcPitch;
     int nDstPitch;
     int nSrcPitchCUDA;
     int nDstPitchCUDA;
     double aBoundingBox[][];
     double nAngle;

     /* 设置显卡,构建上下文 */
     cuRet = cudaSetDevice(  );
     assert( cuRet == cudaSuccess );

     /* 打开文件 */
     pSrcBmp = FreeImage_Load( FIF_BMP, "1.bmp" );
     assert( pSrcBmp != NULL );

     pSrcData = FreeImage_GetBits( pSrcBmp );
     assert( pSrcData != NULL );

     oSrcSize.width = ( int )FreeImage_GetWidth( pSrcBmp );
     oSrcSize.height = ( int )FreeImage_GetHeight( pSrcBmp );
     nSrcPitch = ( int )FreeImage_GetPitch( pSrcBmp );

     oSrcROI.x = oSrcROI.y = ;
     oSrcROI.width = oSrcSize.width;
     oSrcROI.height = oSrcSize.height;

     nAngle = ;

     /* 分配显存 */
     pSrcDataCUDA = nppiMalloc_8u_C1( oSrcSize.width, oSrcSize.height, &nSrcPitchCUDA );
     assert( pSrcDataCUDA != NULL );

     /* 计算旋转后长宽 */
     nppiGetRotateBound( oSrcROI, aBoundingBox, nAngle, ,  );
     oDstSize.width = ( int )ceil( fabs( aBoundingBox[][] - aBoundingBox[][] ) );
     oDstSize.height = ( int )ceil( fabs( aBoundingBox[][] - aBoundingBox[][] ) );

     /* 建目标图 */
     pDstBmp = FreeImage_Allocate( oDstSize.width, oDstSize.height,  );
     assert( pDstBmp != NULL );

     pDstData = FreeImage_GetBits( pDstBmp );

     nDstPitch = ( int )FreeImage_GetPitch( pDstBmp );
     oDstROI.x = oDstROI.y = ;
     oDstROI.width = oDstSize.width;
     oDstROI.height = oDstSize.height;

     /* 分配显存 */
     pDstDataCUDA = nppiMalloc_8u_C1( oDstSize.width, oDstSize.height, &nDstPitchCUDA );
     assert( pDstDataCUDA != NULL );

     cudaMemcpy2D( pSrcDataCUDA, nSrcPitchCUDA, pSrcData, nSrcPitch, oSrcSize.width, oSrcSize.height, cudaMemcpyHostToDevice );
     cudaMemset2D( pDstDataCUDA, nDstPitchCUDA, , oDstSize.width, oDstSize.height );

     /* 处理 */
     nppRet = nppiRotate_8u_C1R( pSrcDataCUDA, oSrcSize, nSrcPitchCUDA, oSrcROI,
                                 pDstDataCUDA, nDstPitchCUDA, oDstROI,
                                 nAngle, - aBoundingBox[][], - aBoundingBox[][], NPPI_INTER_CUBIC );
     assert( nppRet == NPP_NO_ERROR );

     cudaMemcpy2D( pDstData, nDstPitch, pDstDataCUDA, nDstPitchCUDA, oDstSize.width, oDstSize.height, cudaMemcpyDeviceToHost );

     fiRet = FreeImage_Save( FIF_BMP, pDstBmp, "2.bmp" );
     assert( fiRet );

     nppiFree( pSrcDataCUDA );
     nppiFree( pDstDataCUDA );

     cudaDeviceReset();

     FreeImage_Unload( pSrcBmp );
     FreeImage_Unload( pDstBmp );

     return ;
 }