aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

added workable .cu files for stencil & mandelbrot

Browse Source
This commit is contained in:
Evghenii
2013-11-08 10:00:49 +01:00
parent cb7cbec0d5
commit 426afc7377
10 changed files with 645 additions and 160 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
examples_cuda/mandelbrot_tasks3d/mandelbrot_task.ispc
View File

@@ -1,13 +1,10 @@
#ifdef __NVPTX__
#define blockIndex0 blockIndex0()
#define blockIndex1 blockIndex1()
#define vectorWidth warpSize()
#define vectorIndex laneIndex()
#else
#define blockIndex0 taskIndex0
#define blockIndex1 taskIndex1
#define vectorWidth programCount
#define vectorIndex programIndex
#define taskIndex0 blockIndex0()
#define taskIndex1 blockIndex1()
#define taskCount0 blockCount0()
#define taskCount1 blockCount1()
#define programCount warpSize()
#define programIndex laneIndex()
#endif
#if 0
@@ -46,23 +43,25 @@ mandelbrot_scanline(
uniform int xspan, uniform int yspan,
uniform int maxIterations, uniform int output[])
{
const uniform int xstart = blockIndex0 * xspan;
if (taskIndex0 >= taskCount0) return;
if (taskIndex1 >= taskCount1) return;
const uniform int xstart = taskIndex0 * xspan;
const uniform int xend = min(xstart + xspan, width);
const uniform int ystart = blockIndex1 * yspan;
const uniform int ystart = taskIndex1 * yspan;
const uniform int yend = min(ystart + yspan, height);
// assert(xspan >= vectorWidth);
for (uniform int yi = ystart; yi < yend; yi++)
for (uniform int xi = xstart; xi < xend; xi += vectorWidth)
for (uniform int xi = xstart; xi < xend; xi += programCount)
{
const float x = x0 + (xi + vectorIndex) * dx;
const float x = x0 + (xi + programIndex) * dx;
const float y = y0 + yi * dy;
const int res = mandel(x,y,maxIterations);
const int index = yi * width + (xi + vectorIndex);
if (xi + vectorIndex < xend)
const int index = yi * width + (xi + programIndex);
if (xi + programIndex < xend)
output[index] = res;
}
}