void swanBindToTexture1DEx( const char *modname, const char *texname, size_t width, void *ptr, size_t typesize, int flags ) {
CUresult err;
CUtexref cu_texref;
int mode, channels;
// get the module
CUmodule mod = swanGetModule( modname );
// get the texture
err = cuModuleGetTexRef(&cu_texref, mod, texname );
if( err != CUDA_SUCCESS) { error( "swanBindToTexture1D failed -- texture not found" ); }
// now bind
err = cuTexRefSetAddress( NULL, cu_texref, PTR_TO_CUDEVPTR(ptr), width * typesize );
if( err != CUDA_SUCCESS) {
printf("EEERRR = %d\n", err );
error( "swanBindToTexture1D failed -- bind failed" );
}
// does not work for linear memory
/*
if( (flags & TEXTURE_INTERPOLATE) == TEXTURE_INTERPOLATE ) {
err = cuTexRefSetFilterMode( cu_texref, CU_TR_FILTER_MODE_LINEAR );
}
else {
err = cuTexRefSetFilterMode( cu_texref, CU_TR_FILTER_MODE_POINT );
}
if( err != CUDA_SUCCESS) { error( "swanBindToTexture1D failed -- setfiltermode failed" ); }
*/
mode = flags & TEXTURE_TYPE_MASK;
channels = typesize / sizeof(float);
switch( mode ) {
case TEXTURE_FLOAT:
err = cuTexRefSetFormat( cu_texref, CU_AD_FORMAT_FLOAT, channels );
break;
case TEXTURE_INT:
err = cuTexRefSetFormat( cu_texref, CU_AD_FORMAT_SIGNED_INT32, channels );
break;
case TEXTURE_UINT:
err = cuTexRefSetFormat( cu_texref, CU_AD_FORMAT_UNSIGNED_INT32, channels );
break;
default:
error( "swanBinToTexture1D failed -- invalid format" );
}
if( err != CUDA_SUCCESS) {
error( "swanBinToTexture1D failed -- setformat failed" );
}
return;
}
void CudaModule::setTexRef( const std::string& name,
CUdeviceptr ptr,
S64 size,
CUarray_format format,
int numComponents)
{
CUtexref texRef = getTexRef(name);
checkError("cuTexRefSetFormat", cuTexRefSetFormat(texRef, format, numComponents));
checkError("cuTexRefSetAddress", cuTexRefSetAddress(NULL, texRef, ptr, (U32)size));
}
CUresult CuModule::BindLinearTexture(const std::string& name, CuDeviceMem* mem,
CUarray_format format, int numChannels) {
TexBind* texBind;
CUresult result = FindTexRef(name, &texBind);
HANDLE_RESULT();
SetFormat(texBind, format, numChannels);
size_t offset;
result = cuTexRefSetAddress(&offset, texBind->texRef, mem->Handle(),
mem->Size());
HANDLE_RESULT();
texBind->texture.reset();
texBind->mem = mem;
return CUDA_SUCCESS;
}
sparseStatus_t sparseEngine_d::Multiply(sparseMat_t mat, T alpha, T beta,
CUdeviceptr xVec, CUdeviceptr yVec) {
sparseMatrix* m = static_cast<sparseMatrix*>(mat);
Kernel* k;
sparseStatus_t status = LoadKernel(m->prec, &k);
if(SPARSE_STATUS_SUCCESS != status) return status;
// Push the args and select the xVec as a texture
CuCallStack callStack;
callStack.Push(m->outputIndices, m->colIndices, m->sparseValues,
m->tempOutput, m->numGroups);
// Get the size of the xVec elements
PrecTerm precTerms = PrecTerms[m->prec];
size_t offset;
CUresult result = cuTexRefSetAddress(&offset, k->xVec_texture, xVec,
m->width * precTerms.vecSize);
if(CUDA_SUCCESS != result) return SPARSE_STATUS_KERNEL_ERROR;
// Launch the function
uint numBlocks = DivUp(m->numGroups, WarpsPerBlock);
result = k->func[IndexFromVT(m->valuesPerThread)]->Launch(numBlocks, 1,
callStack);
if(CUDA_SUCCESS != result) return SPARSE_STATUS_LAUNCH_ERROR;
// Finalize the vector
int numFinalizeBlocks = DivUp(m->numGroups, WarpsPerBlock);
int useBeta = !IsZero(beta);
callStack.Reset();
callStack.Push(m->tempOutput, m->rowIndices, m->height, yVec, alpha, beta,
useBeta);
result = k->finalize->Launch(numFinalizeBlocks, 1, callStack);
if(CUDA_SUCCESS != result) return SPARSE_STATUS_KERNEL_ERROR;
return SPARSE_STATUS_SUCCESS;
}
void CudaModule::unsetTexRef(const std::string& name)
{
CUtexref texRef = getTexRef(name);
checkError("cuTexRefSetAddress", cuTexRefSetAddress( 0, texRef, 0, 0));
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
CUresult cudastatus = CUDA_SUCCESS;
if (nrhs != 2)
mexErrMsgTxt("Wrong number of arguments");
if (init == 0) {
// Initialize function
//mexLock();
// load GPUmat
gm = gmGetGPUmat();
// load module
CUmodule *drvmod = gmGetModule("examples_texture");
// load float GPU function
CUresult status = cuModuleGetFunction(&drvfunf, *drvmod, "LININTERF");
if (CUDA_SUCCESS != status) {
mexErrMsgTxt("Unable to load user function.");
}
// load double GPU function
status = cuModuleGetFunction(&drvfund, *drvmod, "LININTERD");
if (CUDA_SUCCESS != status) {
mexErrMsgTxt("Unable to load user function.");
}
// load textures defined in module
status = cuModuleGetTexRef(&texf, *drvmod, "texref_f1_a");
if (CUDA_SUCCESS != status) {
mexErrMsgTxt("Unable to load texture.");
}
status = cuModuleGetTexRef(&texd, *drvmod, "texref_d1_a");
if (CUDA_SUCCESS != status) {
mexErrMsgTxt("Unable to load texture.");
}
// no complex function support
init = 1;
}
// mex parameters are:
// 1. IN1. Input array
// 2. IN2. Input indexes array
//IN1 is the input GPU array
GPUtype IN1 = gm->gputype.getGPUtype(prhs[0]);
//IN2 is the input GPU array
GPUtype IN2 = gm->gputype.getGPUtype(prhs[1]);
//OUT is the output GPU array (result)
// Create of the same size of IN1
gpuTYPE_t in1_t = gm->gputype.getType(IN1);
int in1_d = gm->gputype.getNdims(IN1);
const int * in1_s = gm->gputype.getSize(IN1);
int in1_n = gm->gputype.getNumel(IN1);
int in1_b = gm->gputype.getDataSize(IN1);
gpuTYPE_t in2_t = gm->gputype.getType(IN2);
int in2_d = gm->gputype.getNdims(IN2);
const int * in2_s = gm->gputype.getSize(IN2);
int in2_n = gm->gputype.getNumel(IN2);
if ((in1_t==gpuCFLOAT) || (in1_t==gpuCDOUBLE)) {
mexErrMsgTxt("Complex TYPE not supported");
}
if (in1_t != in2_t) {
mexErrMsgTxt("Input arguments must be of the same type");
}
if (in1_n != in2_n) {
mexErrMsgTxt("Input arguments must have the same number of elements");
}
//OUT is the output GPU array (result)
// Create of the same size of IN1
GPUtype OUT = gm->gputype.create(in1_t, in1_d, in1_s, NULL);
// I need the pointers to GPU memory
CUdeviceptr d_IN1 = (CUdeviceptr) (UINTPTR gm->gputype.getGPUptr(IN1));
CUdeviceptr d_IN2 = (CUdeviceptr) (UINTPTR gm->gputype.getGPUptr(IN2));
CUdeviceptr d_OUT = (CUdeviceptr) (UINTPTR gm->gputype.getGPUptr(OUT));
// The GPU kernel depends on the type of input/output
CUfunction drvfun;
CUtexref drvtex;
CUarray_format_enum drvtexformat;
int drvtexnum;
if (in1_t == gpuFLOAT) {
drvfun = drvfunf;
drvtex = texf;
drvtexformat = CU_AD_FORMAT_FLOAT;
drvtexnum = 1;
} else if (in1_t == gpuDOUBLE) {
drvfun = drvfund;
drvtex = texd;
drvtexformat = CU_AD_FORMAT_SIGNED_INT32;
drvtexnum = 2;
}
if (CUDA_SUCCESS != cuTexRefSetFormat(drvtex, drvtexformat, drvtexnum)) {
mexErrMsgTxt("Execution error (texture).");
}
if (CUDA_SUCCESS != cuTexRefSetAddress(NULL, drvtex, UINTPTR d_IN1, in1_n*in1_b)) {
mexErrMsgTxt("Execution error (texture).");
}
if (CUDA_SUCCESS != cuParamSetTexRef(drvfun, CU_PARAM_TR_DEFAULT, drvtex)) {
mexErrMsgTxt("Execution error (texture1).");
}
hostdrv_pars_t gpuprhs[2];
int gpunrhs = 2;
gpuprhs[0] = hostdrv_pars(&d_IN2,sizeof(d_IN2),__alignof(d_IN2));
gpuprhs[1] = hostdrv_pars(&d_OUT,sizeof(d_OUT),__alignof(d_OUT));
int N = in1_n;
hostGPUDRV(drvfun, N, gpunrhs, gpuprhs);
// return result
plhs[0] = gm->gputype.createMxArray(OUT);
}