Setup DimGrid and DimBlock in CUDA -


i doing matrix multiplication in cuda. following setup works:

int tile = 8; dim3 dimgrid((numccolumns - 1)/tile + 1, (numcrows - 1)/tile + 1, 1); dim3 dimblock(tile, tile, 1);

but if use 1 block whole image, returns zero. reason that? assume 1 block can contain whole image ( input 64x64).

dim3 dimgrid(1,1,1); dim3 dimblock(numccolumns, numcrows, 1);

this how call kernel in main function:

matrixmultiply<<<dimgrid, dimblock>>>(devicea, deviceb, devicec,                                         numarows, numacolumns,                                         numbrows, numbcolumns,                                         numcrows, numccolumns);

and kernel:

__global__ void matrixmultiply(float * a, float * b, float * c,                    int numarows, int numacolumns,                    int numbrows, int numbcolumns,                    int numcrows, int numccolumns) {     //@@ insert code implement matrix multiplication here     int row = blockidx.y * blockdim.y + threadidx.y;     int col = blockidx.x * blockdim.x + threadidx.x;      if ((row < numcrows) && (col < numccolumns))     {         float value = 0.0;         (int = 0; < numacolumns; i++)             value += a[row * numacolumns + i] * b[i*numbcolumns + col];         c[row * numccolumns + col] = value;     } }

but if use 1 block whole image, returns zero. reason that?

a cuda threadblock limited maximum of 1024 threads (refer "maximum number of threads per block "). multidimensional threadblock, means product of dimensions must less or equal 1024 (for cc2.x , newer gpus.)

for 64x64 image, not work:

dim3 dimblock(numccolumns, numcrows, 1);

since numccolumns * numcrows greater 1024.

if proper cuda error checking in code, you'll indication of (that kernel launch failing due invalid kernel configuration parameter).


-

Popular posts from this blog

debugging - Reference - What does this error mean in PHP? -

what this? this number of answers warnings, errors , notices might encounter while programming php , have no clue how fix. community wiki, invited participate in adding , maintaining list. why this? questions "headers sent" or "calling member of non-object" pop on stack overflow. root cause of questions same. answers questions typically repeat them , show op line change in his/her particular case. these answers not add value site because apply op's particular code. other users having same error can not read solution out of because localized. sad, because once understood root cause, fixing error trivial. hence, list tries explain solution in general way apply. what should here? if question has been marked duplicate of this, please find error message below , apply fix code. answers contain further links investigate in case shouldn't clear general answer alone. if want contribute, please add "favorite" error message, warning or notic...
Read more

c++ - Why doesn't unordered_set provide an array access operator -

i'm curious why stl container unordered_set , has constant time complexity random access on average, not provide method accessing elements distance first element in container. example: t& unordered_set::operator[](size_t index) { return *(begin() + index); } accessing element "by distance" implies there meaningful way measure distance. trouble std::unordered_set is is, well, unordered . hence, there no meaningful way of explaining "some distance beginning" in non-arbitrary way. if want access distance, copy data vector: std::vector tmp(unordered.begin(), unordered.end());
Read more