/* Returns CL_TRUE on error */
cl_bool findRunSizes(RunSizes* sizes,
const CLInfo* ci,
const DevInfo* di,
const AstronomyParameters* ap,
const IntegralArea* ia,
const CLRequest* clr)
{
WGInfo wgi;
cl_int err;
size_t nWavefrontPerCU;
size_t blockSize; /* Size chunks should be multiples of */
cl_bool forceOneChunk = clr->nonResponsive || di->nonOutput || di->hasGraphicsQOS;
/* I assume this is how this works for 1D limit */
const cl_ulong maxWorkDim = (cl_ulong) di->maxWorkItemSizes[0] * di->maxWorkItemSizes[1] * di->maxWorkItemSizes[2];
const cl_ulong r = (cl_ulong) ia->r_steps;
const cl_ulong mu = (cl_ulong) ia->mu_steps;
sizes->r = ia->r_steps;
sizes->mu = ia->mu_steps;
sizes->nu = ia->nu_steps;
sizes->area = r * mu;
if (r > CL_ULONG_MAX / mu)
{
mw_printf("Integral area overflows cl_ulong\n");
return CL_TRUE;
}
if (di->devType == CL_DEVICE_TYPE_CPU)
{
sizes->nChunk = sizes->nChunkEstimate = 1;
sizes->chunkSize = sizes->effectiveArea = sizes->area;
sizes->extra = 0;
sizes->local[0] = 1;
sizes->global[0] = sizes->area;
return CL_FALSE;
}
err = mwGetWorkGroupInfo(_separationKernel, ci, &wgi);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Failed to get work group info");
return CL_TRUE;
}
if (clr->verbose)
{
mwPrintWorkGroupInfo(&wgi);
}
if (!mwDivisible(wgi.wgs, (size_t) di->warpSize))
{
mw_printf("Kernel reported work group size ("ZU") not a multiple of warp size (%u)\n",
wgi.wgs,
di->warpSize);
return CL_TRUE;
}
/* This should give a good occupancy. If the global size isn't a
* multiple of this bad performance things happen. */
nWavefrontPerCU = wgi.wgs / di->warpSize;
/* Since we don't use any workgroup features, it makes sense to
* use the wavefront size as the workgroup size */
sizes->local[0] = di->warpSize;
/* For maximum efficiency, we want global work sizes to be multiples of
* (warp size) * (number compute units) * (number of warps for good occupancy)
* Then we throw in another factor since we can realistically do more work at once
*/
blockSize = nWavefrontPerCU * di->warpSize * di->maxCompUnits;
{
cl_uint nBlockPerChunk = 1;
sizes->nChunkEstimate = findNChunk(ap, ia, di, clr, &sizes->initialWait);
/* Make a guess appropriate for the hardware. */
/* m * b ~= area / n */
nBlockPerChunk = sizes->area / (sizes->nChunkEstimate * blockSize);
if (nBlockPerChunk == 0)
nBlockPerChunk = 1;
sizes->chunkSize = nBlockPerChunk * blockSize;
}
//sizes->effectiveArea = sizes->chunkSize * mwDivRoundup(sizes->area, sizes->chunkSize);
sizes->effectiveArea = di->warpSize * mwDivRoundup(sizes->area, di->warpSize);
sizes->nChunk = forceOneChunk ? 1 : mwDivRoundup(sizes->effectiveArea, sizes->chunkSize);
sizes->extra = (cl_uint) (sizes->effectiveArea - sizes->area);
if (sizes->nChunk == 1) /* BlockPerChunk factor probably too high or very small workunit, or nonresponsive */
{
/* Like using nBlockPerChunk == 1 */
sizes->effectiveArea = blockSize * mwDivRoundup(sizes->area, blockSize);
sizes->chunkSize = sizes->effectiveArea;
sizes->extra = sizes->effectiveArea - sizes->area;
}
mw_printf("Using a target frequency of %.1f\n"
"Using a block size of "ZU" with "ZU" blocks/chunk\n",
clr->targetFrequency,
blockSize,
sizes->chunkSize / blockSize
);
printPollMode(ci, sizes);
sizes->chunkSize = sizes->effectiveArea / sizes->nChunk;
/* We should be hitting memory size limits before we ever get to this */
if (sizes->chunkSize > maxWorkDim)
{
mw_printf("Warning: Area too large for one chunk (max size = "LLU")\n", maxWorkDim);
while (sizes->chunkSize > maxWorkDim)
{
sizes->nChunk *= 2;
sizes->chunkSize = sizes->effectiveArea / sizes->nChunk;
}
if (!mwDivisible(sizes->chunkSize, sizes->local[0]))
{
mw_printf("FIXME: I'm too lazy to handle very large workunits properly\n");
return CL_TRUE;
}
else if (!mwDivisible(sizes->chunkSize, blockSize))
{
mw_printf("FIXME: Very large workunit potentially slower than it should be\n");
}
}
sizes->global[0] = sizes->chunkSize;
printRunSizes(sizes, ia);
if (sizes->effectiveArea < sizes->area)
{
mw_printf("Effective area less than actual area!\n");
return CL_TRUE;
}
return CL_FALSE;
}
/* Returns CL_TRUE on error */
cl_bool findRunSizes(RunSizes* sizes,
const CLInfo* ci,
const DevInfo* di,
const IntegralArea* ia,
const CLRequest* clr)
{
WGInfo wgi;
cl_int err;
cl_uint groupSize;
size_t i, nMod;
size_t sum = 0;
err = mwGetWorkGroupInfo(ci, &wgi);
if (err != CL_SUCCESS)
{
mwCLWarn("Failed to get work group info", err);
return CL_TRUE;
}
mwPrintWorkGroupInfo(&wgi);
groupSize = mwFindGroupSize(di);
sizes->local[0] = groupSize;
sizes->local[1] = 1;
sizes->nChunkEstimate = chooseNumChunk(ia, clr, di);
sizes->nChunk = sizes->nChunkEstimate;
/* Best for performance.
Be a bit more flexible if chunking.
Using the whole block size seems a bit better when attacking everything at once.
*/
nMod = sizes->nChunk == 1 ? mwBlockSize(di) : groupSize * di->maxCompUnits;
sizes->area = ia->r_steps * ia->mu_steps;
sizes->effectiveArea = nMod * mwDivRoundup(sizes->area, nMod);
sizes->extra = sizes->effectiveArea - sizes->area;
warn("Keeping chunk boundaries as multiples of "ZU"\n", nMod);
if (sizes->effectiveArea / sizes->nChunk < nMod)
{
sizes->nChunk = sizes->effectiveArea / nMod;
warn("Warning: Estimated number of chunks ("ZU") too large. Using "ZU"\n", sizes->nChunkEstimate, sizes->nChunk);
}
if (nMod == 1)
{
/* When nMod = 1 we need to avoid losing pieces at the
* beginning and end; the normal method doesn't quite work. */
while (sizes->nChunk * (sizes->effectiveArea / sizes->nChunk) < sizes->effectiveArea)
{
sizes->nChunk++;
}
warn("Need to use "ZU" chunks to cover area using multiples of 1\n", sizes->nChunk);
}
sizes->chunkBorders = mwCallocA((sizes->nChunk + 1), sizeof(size_t));
for (i = 0; i <= sizes->nChunk; ++i)
{
if (nMod == 1)
{
/* Avoid losing out the 0 border */
sizes->chunkBorders[i] = i * (sizes->effectiveArea / sizes->nChunk);
}
else
{
sizes->chunkBorders[i] = (i * sizes->effectiveArea + sizes->nChunk) / (sizes->nChunk * nMod);
sizes->chunkBorders[i] *= nMod;
}
if (sizes->chunkBorders[i] > sizes->effectiveArea)
sizes->chunkBorders[i] = sizes->effectiveArea;
if (i > 0)
{
sum += sizes->chunkBorders[i] - sizes->chunkBorders[i - 1];
assert(sizes->chunkBorders[i] - sizes->chunkBorders[i - 1] > 0);
}
}
printRunSizes(sizes, ia, clr->verbose);
if (sum != sizes->effectiveArea) /* Assert that the divisions aren't broken */
{
warn("Chunk total does not match: "ZU" != "ZU"\n", sum, sizes->effectiveArea);
free(sizes->chunkBorders);
return CL_TRUE;
}
return CL_FALSE;
}