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

changes

Browse Source
This commit is contained in:
Evghenii
2013-11-18 11:58:19 +01:00
parent 4d278a654e
commit 8d4dd13750
38 changed files with 5481 additions and 4073 deletions
Download Patch File Download Diff File Expand all files Collapse all files

154
examples_cuda/stencil/stencil.ispc
View File

@@ -31,16 +31,6 @@
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifdef __NVPTX__
#warning "emitting DEVICE code"
#define taskIndex blockIndex0()
#define taskCount blockCount0()
#define programIndex laneIndex()
#define programCount warpSize()
#else
#warning "emitting HOST code"
#endif
static inline void
stencil_step(uniform int x0, uniform int x1,
uniform int y0, uniform int y1,
@@ -50,80 +40,59 @@ stencil_step(uniform int x0, uniform int x1,
uniform const double Ain[], uniform double Aout[]) {
const uniform int Nxy = Nx * Ny;
// foreach (z = z0 ... z1, y = y0 ... y1, x = x0 ... x1)
#if 0
#define VER1
#endif
#ifdef VER1
const uniform long x1o = 1;
const uniform long x2o = 2;
const uniform long x3o = 3;
const uniform long y1o = Nx;
const uniform long y2o = Nx*2;
const uniform long y3o = Nx*3;
const uniform long z1o = Nxy;
const uniform long z2o = Nxy*2;
const uniform long z3o = Nxy*3;
#endif
for (uniform int z = z0; z < z1; z++)
for (uniform int y = y0; y < y1; y++)
{
const int index_base = (z * Nxy) + (y * Nx);
for (uniform int xb = x0; xb < x1; xb += programCount)
{
const int x = xb + programIndex;
int index = index_base + x;
#ifndef VER1
foreach (z = z0 ... z1, y = y0 ... y1, x = x0 ... x1)
{
int index = (z * Nxy) + (y * Nx) + x;
#define A_cur(x, y, z) Ain[index + (x) + ((y) * Nx) + ((z) * Nxy)]
#define A_next(x, y, z) Aout[index + (x) + ((y) * Nx) + ((z) * Nxy)]
double div = coef[0] * A_cur(0, 0, 0) +
coef[1] * (A_cur(+1, 0, 0) + A_cur(-1, 0, 0) +
A_cur(0, +1, 0) + A_cur(0, -1, 0) +
A_cur(0, 0, +1) + A_cur(0, 0, -1)) +
coef[2] * (A_cur(+2, 0, 0) + A_cur(-2, 0, 0) +
A_cur(0, +2, 0) + A_cur(0, -2, 0) +
A_cur(0, 0, +2) + A_cur(0, 0, -2)) +
coef[3] * (A_cur(+3, 0, 0) + A_cur(-3, 0, 0) +
A_cur(0, +3, 0) + A_cur(0, -3, 0) +
A_cur(0, 0, +3) + A_cur(0, 0, -3));
#else
#define A_cur(x, y, z) Ain [index + (x) + (y) + (z)]
#define A_next(x, y, z) Aout[index + (x) + (y) + (z)]
double div = coef[0] * A_cur(0, 0, 0) +
coef[1] * (A_cur(+x1o, 0, 0) + A_cur(-x1o, 0, 0) +
A_cur(0, +y1o, 0) + A_cur(0, -y1o, 0) +
A_cur(0, 0, +z1o) + A_cur(0, 0, -z1o)) +
coef[2] * (A_cur(+x2o, 0, 0) + A_cur(-x2o, 0, 0) +
A_cur(0, +y2o, 0) + A_cur(0, -y2o, 0) +
A_cur(0, 0, +z2o) + A_cur(0, 0, -z2o)) +
coef[3] * (A_cur(+x3o, 0, 0) + A_cur(-x3o, 0, 0) +
A_cur(0, +y3o, 0) + A_cur(0, -y3o, 0) +
A_cur(0, 0, +z3o) + A_cur(0, 0, -z3o));
#endif
double div = coef[0] * A_cur(0, 0, 0) +
coef[1] * (A_cur(+1, 0, 0) + A_cur(-1, 0, 0) +
A_cur(0, +1, 0) + A_cur(0, -1, 0) +
A_cur(0, 0, +1) + A_cur(0, 0, -1)) +
coef[2] * (A_cur(+2, 0, 0) + A_cur(-2, 0, 0) +
A_cur(0, +2, 0) + A_cur(0, -2, 0) +
A_cur(0, 0, +2) + A_cur(0, 0, -2)) +
coef[3] * (A_cur(+3, 0, 0) + A_cur(-3, 0, 0) +
A_cur(0, +3, 0) + A_cur(0, -3, 0) +
A_cur(0, 0, +3) + A_cur(0, 0, -3));
if (x < x1)
A_next(0, 0, 0) = 2.0d0 * A_cur(0, 0, 0) - A_next(0, 0, 0) +
vsq[index] * div;
}
}
A_next(0, 0, 0) = 2.0 * A_cur(0, 0, 0) - A_next(0, 0, 0) +
vsq[index] * div;
}
}
#define SPANX 32
#define SPANY 8
#define SPANZ 8
static task void
stencil_step_task(uniform int x0, uniform int x1,
uniform int y0, uniform int y1,
uniform int z0,
uniform int z0, uniform int z1,
uniform int Nx, uniform int Ny, uniform int Nz,
uniform const double coef[4], uniform const double vsq[],
uniform const double Ain[], uniform double Aout[]) {
if(taskIndex >= taskCount) return;
if (taskIndex0 >= taskCount0 ||
taskIndex1 >= taskCount1 ||
taskIndex2 >= taskCount2)
return;
stencil_step(x0, x1, y0, y1, z0+taskIndex, z0+taskIndex+1,
Nx, Ny, Nz, coef, vsq, Ain, Aout);
const uniform int xfirst = x0 + taskIndex0 * SPANX;
const uniform int xlast = min(x1, xfirst + SPANX);
const uniform int yfirst = y0 + taskIndex1 * SPANY;
const uniform int ylast = min(y1, yfirst + SPANY);
const uniform int zfirst = z0 + taskIndex2 * SPANZ;
const uniform int zlast = min(z1, zfirst + SPANZ);
stencil_step(xfirst,xlast, yfirst,ylast, zfirst,zlast,
Nx, Ny, Nz, coef, vsq, Ain, Aout);
}
export void
loop_stencil_ispc_tasks(uniform int t0, uniform int t1,
uniform int x0, uniform int x1,
@@ -134,39 +103,24 @@ loop_stencil_ispc_tasks(uniform int t0, uniform int t1,
uniform const double vsq[],
uniform double Aeven[], uniform double Aodd[])
{
for (uniform int t = t0; t < t1; ++t) {
// Parallelize across cores as well: each task will work on a slice
// of 1 in the z extent of the volume.
if ((t & 1) == 0)
launch[z1-z0] stencil_step_task(x0, x1, y0, y1, z0, Nx, Ny, Nz,
coef, vsq, Aeven, Aodd);
else
launch[z1-z0] stencil_step_task(x0, x1, y0, y1, z0, Nx, Ny, Nz,
coef, vsq, Aodd, Aeven);
#define NB(x,n) (((x)+(n)-1)/(n))
// We need to wait for all of the launched tasks to finish before
// starting the next iteration.
sync;
for (uniform int t = t0; t < t1; ++t)
{
// Parallelize across cores as well: each task will work on a slice
// of 1 in the z extent of the volume.
if ((t & 1) == 0)
launch[NB(z1-z0,SPANZ)][NB(y1-y0,SPANY)][NB(x1-x0,SPANX)]
stencil_step_task(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz,
coef, vsq, Aeven, Aodd);
else
launch[NB(z1-z0,SPANZ)][NB(y1-y0,SPANY)][NB(x1-x0,SPANX)]
stencil_step_task(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz,
coef, vsq, Aodd, Aeven);
// We need to wait for all of the launched tasks to finish before
// starting the next iteration.
sync;
}
}
export void
loop_stencil_ispc(uniform int t0, uniform int t1,
uniform int x0, uniform int x1,
uniform int y0, uniform int y1,
uniform int z0, uniform int z1,
uniform int Nx, uniform int Ny, uniform int Nz,
uniform const double coef[4],
uniform const double vsq[],
uniform double Aeven[], uniform double Aodd[])
{
for (uniform int t = t0; t < t1; ++t) {
if ((t & 1) == 0)
stencil_step(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz, coef, vsq,
Aeven, Aodd);
else
stencil_step(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz, coef, vsq,
Aodd, Aeven);
}
}