void updateTree(icl_buffer *nodelist, icl_buffer *particlesD, icl_buffer *treeD, UINT nParticles, UINT treeHeight, icl_device* dev) {
// compile OpenCL kernels
icl_kernel* update = icl_create_kernel(dev, "kernel/updateTree.cl", "updateTree", KERNEL_BUILD_MACRO, ICL_SOURCE);
UINT nNodes = nParticles * 2 - 1;
#if timing == 1
icl_timer* timer_update = icl_init_timer(ICL_MILLI);
#endif
size_t localSizeB = 256;
size_t globalSizeB = ((nNodes + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_update);
#endif
for(int l = (int)treeHeight; l >= 0; --l)
{
icl_run_kernel(update, 1, &globalSizeB, &localSizeB, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)treeD,
(size_t)0, (void *)particlesD,
sizeof(int), &l,
sizeof(UINT), &nNodes);
}
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_update);
printf("tree update %f\n\n", timer_update->current_time);
#endif
}
UINT buildTree(icl_buffer *nodelist, icl_buffer *particlesD, icl_buffer *treeD, UINT nParticles, icl_device* dev)
{
UINT level = 1;
UINT nNodes = nParticles * 2 - 1;
icl_timer* timer = icl_init_timer(ICL_MILLI);
// void icl_start_timer(icl_timer* timer);
double time = 0;
// overapproximate size of temporal lists
/* struct Node** activelist = (struct Node**)malloc(nParticles * sizeof(struct Node*)); UINT activeN = 0;
struct Node** smalllist = (struct Node**)malloc(nParticles * sizeof(struct Node*)); UINT smallN = 0;
struct Node** nextlist = (struct Node**)malloc(nParticles * sizeof(struct Node*)); UINT nextN = 0;*/
icl_buffer* activelist = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(NodeId) * nParticles);
icl_buffer* smalllist = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(NodeId) * nParticles);
icl_buffer* nextlist = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(NodeId) * nParticles);
icl_buffer* sizes = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(UINT) * 5); // holds the current size of each of 3 buffers:
// 0 activelist
// 1 nodelist
// 2 smalllist
// 3 old max level
// 4 new max level
UINT maxNchunks = ((nParticles / fmin(T, chunk_size)) * 2) -1;
// assert(maxNchunks <= 256 && "adapt implementation"); // TODO allow more than 256 chunks per node
icl_buffer* chunks = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(struct Chunk) * maxNchunks);
icl_buffer* bboxes = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(struct BBox) * maxNchunks);
size_t localSize = 1;
size_t globalSize = 1;
/*
struct Particle* particles = (struct Particle*)malloc(3000 * sizeof(struct Particle));
icl_read_buffer(particlesD, CL_TRUE, sizeof(struct Particle) * 3000, &particles[0], NULL, NULL);
printf("%f %f %f\n", particles[0].pos.x, particles[0].pos.y, particles[0].pos.z);
*/
// compile OpenCL kernels
icl_kernel* init = icl_create_kernel(dev, "kernel/init.cl", "init", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* resetChunks = icl_create_kernel(dev, "kernel/init.cl", "memset_chunks", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* gp2c = icl_create_kernel(dev, "kernel/groupToChunks.cl", "groupParticlesIntoChunks", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* cBBox = icl_create_kernel(dev, "kernel/chunkedBBox.cl", "chunkedBBox", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* bBox = icl_create_kernel(dev, "kernel/bBox.cl", "bBox", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* sln = icl_create_kernel(dev, "kernel/splitLargeNodes.cl", "splitLargeNodes", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* sortP = icl_create_kernel(dev, "kernel/sortParticlesToChilds.cl", "sortParticlesToChilds", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* snf = icl_create_kernel(dev, "kernel/smallNodeFiltering.cl", "smallNodeFiltering", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* pnl = icl_create_kernel(dev, "kernel/packNextlist.cl", "packNextlist", KERNEL_BUILD_MACRO, ICL_SOURCE);
//////////////////////////////////////////////////////////////////////////
icl_kernel* preScan = icl_create_kernel(dev, "kernel/sortP_prescan.cl", "sortP_prescan_chunked", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* postScan = icl_create_kernel(dev, "kernel/sortP_postscan.cl", "sortP_postscan_chunked", KERNEL_BUILD_MACRO, ICL_SOURCE);
segmented_scan_init(nParticles, dev, KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* memset_int_s = icl_create_kernel(dev, "kernel/init.cl", "memset_int_s", KERNEL_BUILD_MACRO, ICL_SOURCE);
// approach with segmented scan
icl_buffer *scan_data = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(int) * nParticles);
icl_buffer *scan_flag = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(int) * nParticles);
icl_buffer* buffered_particles = icl_create_buffer(dev, CL_MEM_READ_WRITE, sizeof(struct Particle) * nParticles);
#if timing == 1
icl_timer* timer_gp2c = icl_init_timer(ICL_MILLI);
icl_timer* timer_cBBox = icl_init_timer(ICL_MILLI);
icl_timer* timer_bBox = icl_init_timer(ICL_MILLI);
icl_timer* timer_sln = icl_init_timer(ICL_MILLI);
icl_timer* timer_sortP = icl_init_timer(ICL_MILLI);
icl_timer* timer_snf = icl_init_timer(ICL_MILLI);
icl_timer* timer_pnl = icl_init_timer(ICL_MILLI);
icl_timer* timer_ran = icl_init_timer(ICL_MILLI);
icl_timer* timer_prescan = icl_init_timer(ICL_MILLI);
icl_timer* timer_scan = icl_init_timer(ICL_MILLI);
icl_timer* timer_postscan = icl_init_timer(ICL_MILLI);
#endif
//add root node to the activelist and initialize size lists
icl_run_kernel(init, 1, &globalSize, &localSize, NULL, NULL, 3,
(size_t)0, (void *)activelist,
(size_t)0, (void *)sizes,
(size_t)0, (void *)particlesD);
UINT activeN = 1;
icl_finish(dev);
// smallest power of 2 bigger or equal to maxxNchnunks
UINT pow2maxNchunks = pow2roundup(maxNchunks);
// processLargeNode
while(activeN != 0)
{
icl_start_timer(timer);
// group triangles into chunks
size_t localSize1 = min(pow2maxNchunks, 256);
#if timing == 1
icl_start_timer(timer_gp2c);
#endif
size_t globalSize1 = ((maxNchunks + localSize1 -1) / localSize1) * localSize1;
// reset chunks
icl_run_kernel(resetChunks, 1, &globalSize1, &localSize1, NULL, NULL, 2,
(size_t)0, (void *)chunks,
sizeof(UINT), &maxNchunks);
globalSize1 = localSize1 * activeN;
// split every node in chunk of chunk_size
icl_run_kernel(gp2c, 1, &globalSize1, &localSize1, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)chunks,
sizeof(UINT), &activeN);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_gp2c);
#endif
// compute per chunk bounding box
size_t localSize2 = chunk_size;
size_t globalSize2 = maxNchunks * chunk_size;
#if timing == 1
icl_start_timer(timer_cBBox);
#endif
icl_run_kernel(cBBox, 1, &globalSize2, &localSize2, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)chunks,
(size_t)0, (void *)bboxes);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_cBBox);
#endif
// compute per node bounding box
size_t localSize3 = min(pow2maxNchunks, 256);
size_t globalSize3 = localSize3 * activeN;
#if timing == 1
icl_start_timer(timer_bBox);
#endif
icl_run_kernel(bBox, 1, &globalSize3, &localSize3, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)bboxes,
sizeof(UINT), &activeN);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_bBox);
#endif
// split large nodes
size_t localSize4 = 256;
size_t globalSize4 = ((activeN + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_sln);
#endif
icl_run_kernel(sln, 1, &globalSize4, &localSize4, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)activelist,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)sizes,
sizeof(UINT), &activeN);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_sln);
#endif
///////////////////////////////////////////////////////////////////////////////
// XXx replaced with segmented scan
//globalSize = (activeN+1) * 256;
#if timing == 1
icl_start_timer(timer_sortP);
#endif
#if DEVICE == ICL_CPU
// sort particles to child nodes
size_t localSize5 = 256;
size_t globalSize5 = ((activeN + 255) / 256) * 256;
icl_run_kernel(sortP, 1, &globalSize5, &localSize5, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)activelist,
(size_t)0, (void *)nextlist,
sizeof(UINT), &activeN
);
#else
// init scan_flag to 1
cl_int initFlag = 1;
size_t np = (size_t)((nParticles + localSize4 -1 ) / localSize4) * localSize4;
icl_run_kernel(memset_int_s, 1, &np, &localSize4, NULL, NULL, 3,
(size_t)0, (void *)scan_flag,
sizeof(cl_int), &initFlag,
sizeof(UINT), &nParticles
);
#if timing == 1
icl_start_timer(timer_prescan);
#endif
// pre-scan fills data0 and data1 with 1 and 0 whenever value < pivot
localSize = chunk_size;
// globalSize = activeN * 256;
globalSize = maxNchunks * chunk_size;
icl_run_kernel(preScan, 1, &globalSize, &localSize, NULL, NULL, 7,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)chunks,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)activelist,
sizeof(UINT), &activeN,
(size_t)0, (void *)scan_data,
(size_t)0, (void *)scan_flag
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_prescan);
#endif
#if timing == 1
icl_start_timer(timer_scan);
#endif
// scan for
scan(scan_data, scan_flag, nParticles);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_scan);
#endif
// copy partially sorted data to the final
icl_copy_buffer(particlesD, buffered_particles, sizeof(struct Particle) * nParticles, NULL, NULL);
// swap(particlesD, buffered_particles);
#if timing == 1
icl_start_timer(timer_postscan);
#endif
localSize = chunk_size;
globalSize = maxNchunks * chunk_size;
icl_run_kernel(postScan, 1, &globalSize, &localSize, NULL, NULL, 7,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)chunks,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)activelist,
sizeof(UINT), &activeN,
(size_t)0, (void *)buffered_particles,
(size_t)0, (void *)scan_data
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_postscan);
#endif
icl_finish(dev);
#endif
///////////////////////////////////////////////////////////////////////////////
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_sortP);
#endif
// small node filtering
size_t localSize6 = 256;
size_t globalSize6 = ((activeN*2 + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_snf);
#endif
icl_run_kernel(snf, 1, &globalSize6, &localSize6, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)smalllist,
(size_t)0, (void *)sizes
//, sizeof(UINT), &nParticles
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_snf);
#endif
// packing of nextlist
size_t localSize7 = 1;
size_t globalSize7 = 1;
#if timing == 1
icl_start_timer(timer_pnl);
#endif
icl_run_kernel(pnl, 1, &globalSize7, &localSize7, NULL, NULL, 2,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)sizes
);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_pnl);
#endif
// swap nextlist and activelist
swap(&nextlist, &activelist);
#if timing == 1
icl_start_timer(timer_ran);
#endif
// read size of next activelist set in kernel
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT), &activeN, NULL, NULL);
icl_finish(dev);
#if timing == 1
icl_stop_timer(timer_ran);
#endif
++level;
//printf("%d: ActiveN %d\n", level, activeN);
time = icl_stop_timer(timer);
}
icl_release_kernel(init);
icl_release_kernel(gp2c);
icl_release_kernel(cBBox);
icl_release_kernel(bBox);
icl_release_kernel(sln);
icl_release_kernel(sortP);
icl_release_kernel(snf);
icl_release_kernel(pnl);
//////////////////////////////////////////////////////////////////////////
icl_release_kernel(preScan);
icl_release_kernel(postScan);
segmented_scan_release();
icl_release_buffers(3, scan_data, scan_flag, buffered_particles);
#if timing == 1
printf("gp2c %f\ncBBox %f\nbBox %f\nsln %f\nsortP %f\nsnf %f\npnl %f\nran %f\n\n",
timer_gp2c->current_time,
timer_cBBox->current_time,
timer_bBox->current_time,
timer_sln->current_time,
timer_sortP->current_time,
timer_snf->current_time,
timer_pnl->current_time,
timer_ran->current_time);
icl_release_timer(timer_gp2c);
icl_release_timer(timer_cBBox);
icl_release_timer(timer_bBox);
icl_release_timer(timer_sln);
icl_release_timer(timer_sortP);
icl_release_timer(timer_snf);
icl_release_timer(timer_pnl);
printf("prescan %f\nscan %f\npostscan %f\n\n", timer_prescan->current_time, timer_scan->current_time, timer_postscan->current_time);
icl_release_timer(timer_prescan);
icl_release_timer(timer_scan);
icl_release_timer(timer_postscan);
#endif
/*
icl_read_buffer(nodelist, CL_TRUE, sizeof(struct Node) * 6000, tree->nodelist, NULL, NULL);
printf("node: %d, left %d, right %d", tree->nodelist[0].particlesHigh - tree->nodelist[0].particlesLow,
tree->nodelist[1].particlesHigh - tree->nodelist[1].particlesLow, tree->nodelist[2].particlesHigh - tree->nodelist[2].particlesLow);
printBox(tree->nodelist[49].bounding_box);
printBox(tree->nodelist[53].bounding_box);
printBox(tree->nodelist[54].bounding_box);
for(int i = 0; i < 6000; ++i)
if(tree->nodelist[i].bounding_box.box[0].x != 0.0)
printBox(tree->nodelist[i].bounding_box);
*/
//small nodes stage
// preprocessSmallNodes(smalllist);
icl_release_buffers(3, activelist, chunks, bboxes);
icl_kernel* sasl = icl_create_kernel(dev, "kernel/swapActiveAndSmalllist.cl", "swapActiveAndSmalllist", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* ssn = icl_create_kernel(dev, "kernel/splitSmallNodes.cl", "splitSmallNodes", KERNEL_BUILD_MACRO, ICL_SOURCE);
#if timing == 1
icl_timer* timer_ssn = icl_init_timer(ICL_MILLI);
icl_timer* timer_sasl = icl_init_timer(ICL_MILLI);
icl_timer* timer_rsn = icl_init_timer(ICL_MILLI);
#endif
size_t localSize8 = 1;
size_t globalSize8 = 1;
UINT setMaxLevel = 0;
icl_run_kernel(sasl, 1, &globalSize8, &localSize8, NULL, NULL, 2,
(size_t)0, (void *)sizes,
sizeof(UINT), &level);
// get number of small nodes
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT), &activeN, NULL, NULL);
while(activeN != 0)
{
icl_start_timer(timer);
// compute SVH and determine the split plane
size_t localSize9 = 256;
size_t globalSize9 = ((activeN + 255) / 256) * 256;
#if timing == 1
icl_start_timer(timer_ssn);
#endif
icl_run_kernel(ssn, 1, &globalSize9, &localSize9, NULL, NULL, 5,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)smalllist,
(size_t)0, (void *)nextlist,
(size_t)0, (void *)particlesD,
(size_t)0, (void *)sizes);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_ssn);
#endif
size_t localSizeA = 1;
size_t globalSizeA = 1;
#if timing == 1
icl_start_timer(timer_sasl);
#endif
icl_run_kernel(sasl, 1, &globalSizeA, &localSizeA, NULL, NULL, 2,
(size_t)0, (void *)sizes,
sizeof(UINT), &setMaxLevel);
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_sasl);
#endif
swap(&nextlist, &smalllist);
// read size of next activelist set in kernel
#if timing == 1
icl_start_timer(timer_rsn);
#endif
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT), &activeN, NULL, NULL);
//printf("small size %d\n", activeN);
icl_finish(dev);
#if timing == 1
icl_stop_timer(timer_rsn);
#endif
time = icl_stop_timer(timer);
}
icl_release_buffer(smalllist);
icl_release_buffer(nextlist);
icl_release_kernel(sasl);
icl_release_kernel(ssn);
#if timing == 1
printf("ssn %f\nsasl %f\nrsn %f\n\n", timer_ssn->current_time, timer_sasl->current_time, timer_rsn->current_time);
icl_release_timer(timer_ssn);
icl_release_timer(timer_sasl);
icl_release_timer(timer_rsn);
#endif
UINT s[5];
icl_read_buffer(sizes, CL_TRUE, sizeof(UINT) * 5, &s, NULL, NULL);
icl_release_buffer(sizes);
icl_kernel* upPass = icl_create_kernel(dev, "kernel/upPass.cl", "upPass", KERNEL_BUILD_MACRO, ICL_SOURCE);
icl_kernel* downPass = icl_create_kernel(dev, "kernel/kdDownPass.cl", "downPass", KERNEL_BUILD_MACRO, ICL_SOURCE);
#if timing == 1
icl_timer* timer_upPass = icl_init_timer(ICL_MILLI);
icl_timer* timer_downPass = icl_init_timer(ICL_MILLI);
icl_timer* timer_rt = icl_init_timer(ICL_MILLI);
#endif
UINT treeHeight = s[4];
printf("Tree height: %d\n", treeHeight);
size_t localSizeB = 256;
size_t globalSizeB = ((nNodes + 255) / 256) * 256;
icl_start_timer(timer);
#if timing == 1
icl_start_timer(timer_upPass);
#endif
for(int l = (int)treeHeight; l >= 0; --l)
{
icl_run_kernel(upPass, 1, &globalSizeB, &localSizeB, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)particlesD,
sizeof(int), &l,
sizeof(UINT), &nNodes);
}
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_upPass);
icl_start_timer(timer_downPass);
#endif
for(UINT l = 0; l <= treeHeight; ++l)
{
icl_run_kernel(downPass, 1, &globalSizeB, &localSizeB, NULL, NULL, 4,
(size_t)0, (void *)nodelist,
(size_t)0, (void *)treeD,
sizeof(UINT), &l,
sizeof(UINT), &nNodes);
}
icl_finish(dev);
#if timing == 1
icl_stop_timer(timer_downPass);
#endif
time = icl_stop_timer(timer);
icl_release_kernel(upPass);
icl_release_kernel(downPass);
#if timing == 1
icl_start_timer(timer_rt);
#endif
#if timing == 1
icl_finish(dev);
icl_stop_timer(timer_rt);
printf("upPass %f\ndownPass %f\nread Tree %f\n\n", timer_upPass->current_time, timer_downPass->current_time, timer_rt->current_time);
icl_release_timer(timer_upPass);
icl_release_timer(timer_downPass);
icl_release_timer(timer_rt);
#endif
// struct Node* kdTree = (struct Node*)malloc(sizeof(struct Node) * nNodes);
// icl_read_buffer(treeD, CL_TRUE, sizeof(struct Node) * nNodes, kdTree, NULL, NULL);
// printf("%d", tree->nodelist[0].left_child);
printf("\nTime: %f\n", time);
icl_release_timer(timer);
return treeHeight;
}
void scan(icl_buffer *data, icl_buffer *flag, UINT n) {
// use size for actual n, not overapproximation as in allocation
UINT numWorkGroups = ((n + wx - 1) / wx);
UINT sizeScanBuff = ((numWorkGroups + wx -1) / wx) * wx;
size_t gx = numWorkGroups * wx;
INT init = -1;
#if TIMING
clFinish(dev->queue);
icl_start_timer(timer);
#endif
icl_run_kernel(perBlockScanByKey, 1, &gx, &wx, NULL, perBlockScanEvent, 8,
(size_t)0, (void *)flag,
(size_t)0, (void *)data,
sizeof(INT), &init,
sizeof(UINT), &n,
sizeof(UINT) * wx, NULL,
sizeof(INT) * wx, NULL,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)preSumArray);
/*
V_OPENCL( kernels[0].setArg( 0, firstKey.getBuffer()), "Error setArg kernels[ 0 ]" ); // Input keys
V_OPENCL( kernels[0].setArg( 1, firstValue.getBuffer()),"Error setArg kernels[ 0 ]" ); // Input buffer
V_OPENCL( kernels[0].setArg( 2, result.getBuffer( ) ), "Error setArg kernels[ 0 ]" ); // Output buffer
V_OPENCL( kernels[0].setArg( 3, init ), "Error setArg kernels[ 0 ]" ); // Initial value exclusive
V_OPENCL( kernels[0].setArg( 4, numElements ), "Error setArg kernels[ 0 ]" ); // Size of scratch buffer
V_OPENCL( kernels[0].setArg( 5, ldsKeySize, NULL ), "Error setArg kernels[ 0 ]" ); // Scratch buffer
V_OPENCL( kernels[0].setArg( 6, ldsValueSize, NULL ), "Error setArg kernels[ 0 ]" ); // Scratch buffer
V_OPENCL( kernels[0].setArg( 7, *binaryPredicateBuffer),"Error setArg kernels[ 0 ]" ); // User provided functor
V_OPENCL( kernels[0].setArg( 8, *binaryFunctionBuffer ),"Error setArg kernels[ 0 ]" ); // User provided functor
V_OPENCL( kernels[0].setArg( 9, *keySumArray ), "Error setArg kernels[ 0 ]" ); // Output per block sum
V_OPENCL( kernels[0].setArg(10, *preSumArray ), "Error setArg kernels[ 0 ]" ); // Output per block sum
V_OPENCL( kernels[0].setArg(11, doExclusiveScan ), "Error setArg kernels[ 0 ]" ); // Exclusive scan?
*/
UINT workPerThread = sizeScanBuff / wx;
icl_run_kernel(intraBlockInclusiveScanByKey, 1, &wx, &wx, NULL, intraBlockEvent, 7,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)preSumArray,
(size_t)0, (void *)postSumArray,
sizeof(UINT), &numWorkGroups,
sizeof(UINT) * wx, NULL,
sizeof(INT) * wx, NULL,
sizeof(UINT), &workPerThread);
/*
INT* output = (UINT*)malloc(n * sizeof(INT));
icl_read_buffer(keySumArray, CL_TRUE, n * sizeof(INT), output, NULL, NULL);
for(int i = 0; i < n; ++i)
printf("%d ", output[i]);
printf("\n"); */
/*
V_OPENCL( kernels[1].setArg( 0, *keySumArray ), "Error setArg kernels[ 1 ]" ); // Input keys
V_OPENCL( kernels[1].setArg( 1, *preSumArray ), "Error setArg kernels[ 1 ]" ); // Input buffer
V_OPENCL( kernels[1].setArg( 2, *postSumArray ), "Error setArg kernels[ 1 ]" ); // Output buffer
V_OPENCL( kernels[1].setArg( 3, numWorkGroupsK0 ), "Error setArg kernels[ 1 ]" ); // Size of scratch buffer
V_OPENCL( kernels[1].setArg( 4, ldsKeySize, NULL ), "Error setArg kernels[ 1 ]" ); // Scratch buffer
V_OPENCL( kernels[1].setArg( 5, ldsValueSize, NULL ), "Error setArg kernels[ 1 ]" ); // Scratch buffer
V_OPENCL( kernels[1].setArg( 6, workPerThread ), "Error setArg kernels[ 1 ]" ); // User provided functor
V_OPENCL( kernels[1].setArg( 7, *binaryPredicateBuffer ),"Error setArg kernels[ 1 ]" ); // User provided functor
V_OPENCL( kernels[1].setArg( 8, *binaryFunctionBuffer ),"Error setArg kernels[ 1 ]" ); // User provided functor
*/
icl_run_kernel(perBlockAdditionByKey, 1, &gx, &wx, NULL, perBlockAdditionEvent, 5,
(size_t)0, (void *)keySumArray,
(size_t)0, (void *)postSumArray,
(size_t)0, (void *)flag,
(size_t)0, (void *)data,
sizeof(UINT), &n);
/*
V_OPENCL( kernels[2].setArg( 0, *keySumArray ), "Error setArg kernels[ 2 ]" ); // Input buffer
V_OPENCL( kernels[2].setArg( 1, *postSumArray ), "Error setArg kernels[ 2 ]" ); // Input buffer
V_OPENCL( kernels[2].setArg( 2, firstKey.getBuffer()), "Error setArg kernels[ 2 ]" ); // Output buffer
V_OPENCL( kernels[2].setArg( 3, result.getBuffer()), "Error setArg kernels[ 2 ]" ); // Output buffer
V_OPENCL( kernels[2].setArg( 4, numElements ), "Error setArg kernels[ 2 ]" ); // Size of scratch buffer
V_OPENCL( kernels[2].setArg( 5, *binaryPredicateBuffer ),"Error setArg kernels[ 2 ]" ); // User provided functor
V_OPENCL( kernels[2].setArg( 6, *binaryFunctionBuffer ),"Error setArg kernels[ 2 ]" ); // User provided functor
*/
#if TIMING
clFinish(dev->queue);
icl_stop_timer(timer);
perBlockScanTime += icl_profile_event(perBlockScanEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
intraBlockTime += icl_profile_event(intraBlockEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
perBlockAdditionTime += icl_profile_event(perBlockAdditionEvent, MEASURE_START, ICL_FINISHED, ICL_MILLI);
#endif
}
int main(int argc, char* argv[]) {
int size = 1000;
int* input1 = (int*)malloc(sizeof(int) * size);
int* input2 = (int*) malloc(sizeof(int) * size);
int* output = (int *)malloc(sizeof(int) * size);
for(int i=0; i < size; ++i) {
input1[i] = i;
input2[i] = 1;
}
#ifndef INSIEME
icl_timer* time1 = icl_init_timer(ICL_SEC);
icl_start_timer(time1);
#endif
icl_init_devices(ICL_CPU);
#ifndef INSIEME
printf("TIME for initialization: %f\n", icl_stop_timer(time1));
#endif
if (icl_get_num_devices() != 0) {
icl_device* dev = icl_get_device(0);
icl_print_device_short_info(dev);
icl_kernel* kernel = icl_create_kernel(dev, "vec_mul.cl", "vec_mul", "", ICL_SOURCE);
icl_buffer* buf_input1 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(int) * size);
icl_buffer* buf_input2 = icl_create_buffer(dev, CL_MEM_READ_ONLY, sizeof(int) * size);
icl_buffer* buf_output = icl_create_buffer(dev, CL_MEM_WRITE_ONLY, sizeof(int) * size);
icl_event* wb1 = icl_create_event();
icl_event* wb2 = icl_create_event();
icl_event* rb = icl_create_event();
icl_write_buffer(buf_input1, CL_FALSE, sizeof(int) * size, &input1[0], NULL, wb1);
icl_write_buffer(buf_input2, CL_FALSE, sizeof(int) * size, &input2[0], NULL, wb2);
size_t szLocalWorkSize = 256;
float multiplier = size/(float)szLocalWorkSize;
if(multiplier > (int)multiplier)
multiplier += 1;
size_t szGlobalWorkSize = (int)multiplier * szLocalWorkSize;
icl_event* rk = icl_create_event();
icl_event* wb_all = icl_create_event_list(2, wb1, wb2);
icl_run_kernel(kernel, 1, &szGlobalWorkSize, &szLocalWorkSize, wb_all, rk, 4,
(size_t)0, (void *)buf_input1,
(size_t)0, (void *)buf_input2,
(size_t)0, (void *)buf_output,
sizeof(cl_int), (void *)&size);
icl_read_buffer(buf_output, CL_TRUE, sizeof(int) * size, &output[0], rk, rb);
printf("Time wb1 %f\n", icl_profile_event(wb1, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time wb2 %f\n", icl_profile_event(wb2, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time rk %f\n", icl_profile_event(rk, ICL_STARTED, ICL_FINISHED, ICL_SEC));
printf("Time rb %f\n", icl_profile_event(rb, ICL_STARTED, ICL_FINISHED, ICL_SEC));
icl_release_events(5, wb1, wb2, wb_all, rk, rb);
icl_release_buffers(3, buf_input1, buf_input2, buf_output);
icl_release_kernel(kernel);
}
#ifndef INSIEME
icl_restart_timer(time1);
#endif
icl_release_devices();
#ifndef INSIEME
printf("TIME for releasing the devices: %f\n", icl_stop_timer(time1));
icl_release_timer(time1);
#endif
// CHECK for output
printf("======================\n= Vector Mul Done\n");
unsigned int check = 1;
for(unsigned int i = 0; i < size; ++i) {
if(output[i] != i*size) {
check = 0;
printf("= fail at %d, expected %d / actual %d", i, i*3/2, output[i]);
break;
}
}
printf("= result check: %s\n======================\n", check ? "OK" : "FAIL");
free(input1);
free(input2);
free(output);
}
