Base version for BG/Q that is actually working (with disapointing per…

…formance)
Meinersbur · Nov 12, 2012 · 4c2dc40 · 4c2dc40
1 parent 8b37a8d
commit 4c2dc40
Show file tree

Hide file tree

Showing 16 changed files with 315 additions and 273 deletions.
diff --git a/.cproject b/.cproject
@@ -26,6 +26,7 @@
 									<listOptionValue builtIn="false" value="BGQ_QPX=1"/>
 									<listOptionValue builtIn="false" value="SPI=1"/>
 									<listOptionValue builtIn="false" value="MPI=1"/>
+									<listOptionValue builtIn="false" value="PAPI=1"/>
 								</option>
 								<option id="gnu.cpp.compiler.option.include.paths.99280228" name="Include paths (-I)" superClass="gnu.cpp.compiler.option.include.paths" valueType="includePath">
 									<listOptionValue builtIn="false" value="/usr/include/openmpi"/>
@@ -38,6 +39,7 @@
 									<listOptionValue builtIn="false" value="BGQ_QPX=1"/>
 									<listOptionValue builtIn="false" value="SPI=1"/>
 									<listOptionValue builtIn="false" value="MPI=1"/>
+									<listOptionValue builtIn="false" value="PAPI=1"/>
 								</option>
 								<option id="gnu.c.compiler.option.include.paths.318363256" name="Include paths (-I)" superClass="gnu.c.compiler.option.include.paths" valueType="includePath">
 									<listOptionValue builtIn="false" value="/usr/include/openmpi"/>

diff --git a/DirectPut.c b/DirectPut.c
@@ -410,7 +410,7 @@ int msg_InjFifoInit ( msg_InjFifoHandle_t *injFifoHandlePtr,
 	return rc;
       }
 
-      if (g_proc_id==0) printf( "HW freespace=%lx\n", MUSPI_getHwFreeSpace(MUSPI_IdToInjFifo(fifoIds[i],&info->subgroup[subgroupId])) );
+      //if (g_proc_id==0) printf( "HW freespace=%lx\n", MUSPI_getHwFreeSpace(MUSPI_IdToInjFifo(fifoIds[i],&info->subgroup[subgroupId])) );
     }
 
     // Activate the fifos.

diff --git a/bgq/bgq_HoppingMatrix.c b/bgq/bgq_HoppingMatrix.c
@@ -215,12 +215,11 @@ static void bgq_HoppingMatrix_worker_body(void *argptr, size_t tid, size_t threa
 }
 
 
-
-static inline void bgq_HoppingMatrix_worker_readFulllayout(void * restrict arg, size_t tid, size_t threads, bool kamul, bool readFulllayout) {
+static inline void bgq_HoppingMatrix_worker(void * restrict arg, size_t tid, size_t threads, bool kamul, bool readFulllayout) {
 	bgq_HoppingMatrix_workload *work = arg;
 	bool isOdd = work->isOdd_src;
-	bgq_weylfield_controlblock *spinorfield = work->spinorfield;
-	bgq_weylfield_controlblock *targetfield = work->targetfield;
+	bgq_weylfield_controlblock * restrict spinorfield = work->spinorfield;
+	bgq_weylfield_controlblock * restrict targetfield = work->targetfield;
 	ucoord ic_begin = work->ic_begin;
 	ucoord ic_end = work->ic_end;
 	bool noprefetchstream = work->noprefetchstream;
@@ -260,14 +259,27 @@ static inline void bgq_HoppingMatrix_worker_readFulllayout(void * restrict arg,
 
 		bgq_gaugesite *gaugesite = &g_bgq_gaugefield_fromCollapsed[isOdd][ic];
 		bgq_weyl_ptr_t *destptrs = &targetfield->sendptr[ic];
+#if 0
+		bgq_weyl_ptr_t destptrsx = {
+				&targetfield->sec_collapsed->d[TUP],
+				&targetfield->sec_collapsed->d[TDOWN],
+				&targetfield->sec_collapsed->d[XUP],
+				&targetfield->sec_collapsed->d[XDOWN],
+				&targetfield->sec_collapsed->d[YUP],
+				&targetfield->sec_collapsed->d[YDOWN],
+				&targetfield->sec_collapsed->d[ZUP],
+				&targetfield->sec_collapsed->d[ZDOWN]
+		};
+		bgq_weyl_ptr_t * restrict destptrs = &destptrsx;
+#endif
 
 		//TODO: prefetching
 		//TODO: Check inlining
 		bgq_su3_spinor_decl(spinor);
 		if (readFulllayout) {
 			bgq_spinorsite *spinorsite = &spinorfield->sec_fullspinor[ic];
 			assert(spinorsite->s[1][0][0]!=0);
-			bgq_su3_spinor_prefetch_double(&spinorfield->sec_fullspinor[ic+1]); // TODO: This prefetch is too early
+			//bgq_su3_spinor_prefetch_double(&spinorfield->sec_fullspinor[ic+1]); // TODO: This prefetch is too early
 			bgq_HoppingMatrix_loadFulllayout(spinor, spinorsite, t1, t2, x, y, z);
 		} else {
 			bgq_weylsite *weylsite = &spinorfield->sec_collapsed[ic];
@@ -279,17 +291,16 @@ static inline void bgq_HoppingMatrix_worker_readFulllayout(void * restrict arg,
 }
 
 static void bgq_HoppingMatrix_nokamul_worker_readFulllayout(void *arg, size_t tid, size_t threads) {
-	bgq_HoppingMatrix_worker_readFulllayout(arg,tid,threads,false,true);
+	bgq_HoppingMatrix_worker(arg,tid,threads,false,true);
 }
-
 static void bgq_HoppingMatrix_kamul_worker_readFulllayout(void *arg, size_t tid, size_t threads) {
-	bgq_HoppingMatrix_worker_readFulllayout(arg,tid,threads,true,true);
+	bgq_HoppingMatrix_worker(arg,tid,threads,true,true);
 }
 static void bgq_HoppingMatrix_nokamul_worker_readWeyllayout(void *arg, size_t tid, size_t threads) {
-	bgq_HoppingMatrix_worker_readFulllayout(arg,tid,threads,false,false);
+	bgq_HoppingMatrix_worker(arg,tid,threads,false,false);
 }
 static void bgq_HoppingMatrix_kamul_worker_readWeyllayout(void *arg, size_t tid, size_t threads) {
-	bgq_HoppingMatrix_worker_readFulllayout(arg,tid,threads,true,false);
+	bgq_HoppingMatrix_worker(arg,tid,threads,true,false);
 }
 
 
@@ -764,7 +775,7 @@ void bgq_HoppingMatrix(bool isOdd, bgq_weylfield_controlblock *targetfield, bgq_
 				bgq_master_call(&bgq_HoppingMatrix_kamul_worker_readWeyllayout, &work_body);
 		}
 
-		if (!COMM_X) {
+		if (!COMM_T) {
 			// Copy the data from HALO_T into the required locations
 			bgq_master_sync();
 			static bgq_work_datamove work_datamovet;

diff --git a/bgq/bgq_HoppingMatrix.h b/bgq/bgq_HoppingMatrix.h
@@ -440,38 +440,39 @@ EXTERN_INLINE void bgq_HoppingMatrix_compute_storeWeyllayout_alldir_raw(bgq_weyl
 	//TODO: There is some spilling that can proberbly be avoided
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[TDOWN]);
-	bgq_prefetch(&targetptrs->d[TDOWN]);
+	//bgq_prefetch(&targetptrs->d[TDOWN]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_tup(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[XUP]);
-	bgq_prefetch(&targetptrs->d[XUP]);
+	//bgq_prefetch(&targetptrs->d[XUP]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_tdown(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[XDOWN]);
-	bgq_prefetch(&targetptrs->d[XDOWN]);
+	//bgq_prefetch(&targetptrs->d[XDOWN]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_xup(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[YUP]);
-	bgq_prefetch(&targetptrs->d[YUP]);
+	//bgq_prefetch(&targetptrs->d[YUP]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_xdown(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[YDOWN]);
-	bgq_prefetch(&targetptrs->d[TDOWN]);
+	//bgq_prefetch(&targetptrs->d[TDOWN]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_yup(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[ZUP]);
-	bgq_prefetch(&targetptrs->d[ZUP]);
+	//bgq_prefetch(&targetptrs->d[ZUP]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_ydown(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[ZDOWN]);
-	bgq_prefetch(&targetptrs->d[ZDOWN]);
+	//bgq_prefetch(&targetptrs->d[ZDOWN]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_zup(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 
 	bgq_su3_matrix_prefetch_double(&gaugesite->su3[TUP]);
-	bgq_prefetch(&targetptrs->d[TUP]);
+	//bgq_prefetch(&targetptrs->d[TUP]);
 	bgq_HoppingMatrix_compute_storeWeyllayout_zdown(targetptrs, gaugesite, spinor, t1, t2, x, y, z, qka0,qka1,qka2,qka3,kamul);
 }
 
+
 #define bgq_HoppingMatrix_compute_storeWeyllayout(targetptrs,gaugesite,spinor,t1,t2,x,y,z,kamul) bgq_HoppingMatrix_compute_storeWeyllayout_raw(targetptrs,gaugesite,bgq_su3_spinor_vars(spinor),t1,t2,x,y,z,kamul)
 EXTERN_INLINE void bgq_HoppingMatrix_compute_storeWeyllayout_raw(bgq_weyl_ptr_t *targetptrs, bgq_gaugesite *gaugesite, bgq_su3_spinor_params(spinor), ucoord t1, ucoord t2, ucoord x, ucoord y, ucoord z, bool kamul)
 {

diff --git a/bgq/bgq_comm.c b/bgq/bgq_comm.c
@@ -1606,20 +1606,19 @@ static unsigned mySpirank;
 static inline unsigned bgq_abcde2spirank(Personality_t *pers, uint8_t a, uint8_t b,uint8_t c,uint8_t d,uint8_t e) {
 	assert(pers);
 	torus_t tdims = {
-				pers->Network_Config.Anodes,
-				pers->Network_Config.Bnodes,
-				pers->Network_Config.Cnodes,
-				pers->Network_Config.Dnodes,
-				pers->Network_Config.Enodes
-		};
-	torus_t dims =
-	    {
-	      pers->Network_Config.Anodes,
-	      pers->Network_Config.Bnodes,
-	      pers->Network_Config.Cnodes,
-	      pers->Network_Config.Dnodes,
-	      pers->Network_Config.Enodes
-	    };
+		pers->Network_Config.Anodes,
+		pers->Network_Config.Bnodes,
+		pers->Network_Config.Cnodes,
+		pers->Network_Config.Dnodes,
+		pers->Network_Config.Enodes
+	};
+	torus_t dims = {
+	  pers->Network_Config.Anodes,
+	  pers->Network_Config.Bnodes,
+	  pers->Network_Config.Cnodes,
+	  pers->Network_Config.Dnodes,
+	  pers->Network_Config.Enodes
+	};
 
 	unsigned numNodes = tdims.a * tdims.b * tdims.c * tdims.d * tdims.e;
 	unsigned result = ((((a)*dims.b + b)*dims.c + c)*dims.d + d)*dims.e + e;
@@ -1671,7 +1670,7 @@ static void setup_destinations(Personality_t *pers) {
 		MUSPI_SetUpDestination(&nb2dest[cd].dest, nb.a, nb.b, nb.c, nb.d, nb.e);
 		nb2dest[cd].hintsABCD = 0;
 		nb2dest[cd].hintsE = 0;
-		printf("node %d: %d(%d,%d,%d,%d,%d)-%d->%d(%d,%d,%d,%d,%d)\n", g_proc_id, mySpirank, tcoords.a, tcoords.b, tcoords.c, tcoords.e, tcoords.d, cd, nbrank, nb.a, nb.b, nb.c, nb.d, nb.e);
+		//printf("node %d: %d(%d,%d,%d,%d,%d)-%d->%d(%d,%d,%d,%d,%d)\n", g_proc_id, mySpirank, tcoords.a, tcoords.b, tcoords.c, tcoords.e, tcoords.d, cd, nbrank, nb.a, nb.b, nb.c, nb.d, nb.e);
 	}
 }
 
@@ -1835,7 +1834,7 @@ void bgq_comm_spi_init(void) {
 		assert((roffsets[cd] + secsize) <= (bgq_weyl_section_offset(sec_recv_end) - bgq_weyl_section_offset(sec_recv_begin)));
 		totalMessageSize += secsize;
 
-		master_print("SPI %llu: d=%llu msize=%zu soffset=%zu d_dst=%llu roffset=%zu\n", cd, d_src, messageSizes[commdir], soffsets[commdir], d_dst, roffsets[cd]);
+		//master_print("SPI %llu: d=%llu msize=%zu soffset=%zu d_dst=%llu roffset=%zu\n", cd, d_src, messageSizes[commdir], soffsets[commdir], d_dst, roffsets[cd]);
 	}
 	assert(totalMessageSize == bgq_weyl_section_offset(sec_recv_end) - bgq_weyl_section_offset(sec_recv_begin));
 
@@ -1910,7 +1909,7 @@ void bgq_comm_spi_init(void) {
 //TODO: inline?
 void bgq_comm_recv(bool nospi) {
 	assert(omp_get_thread_num()==0);
-	master_print("Comm Receiving...\n");
+	//master_print("Comm Receiving...\n");
 #ifdef SPI
 	if (!nospi) {
 	    // reset the recv counter
@@ -1925,7 +1924,7 @@ void bgq_comm_recv(bool nospi) {
 
 void bgq_comm_send(bool nospi) {
 	assert(omp_get_thread_num()==0);
-	master_print("Comm Sending...\n");
+	//master_print("Comm Sending...\n");
 #ifdef SPI
 	if (!nospi) {
 		// make sure everybody has reset recvCounter
@@ -1947,7 +1946,7 @@ void bgq_comm_send(bool nospi) {
 
 void bgq_comm_wait(bool nospi) {
 	assert(omp_get_thread_num()==0);
-	master_print("Comm Waiting...\n");
+	//master_print("Comm Waiting...\n");
 
 	uint64_t ppc32 = mfspr(SPRN_PPR32);
 	ThreadPriority_Low(); // If there is some other work to be done on this node, give it priority
@@ -1958,17 +1957,17 @@ void bgq_comm_wait(bool nospi) {
 		uint64_t startTime = 0;
 
 		// Wait for all data is received
-		 printf("node %d: %llu bytes to be received\n", g_proc_id, totalMessageSize);
+		 //printf("node %d: %llu bytes to be received\n", g_proc_id, totalMessageSize);
 		 while(recvCounter > 0) {
 			 // Check range of pending bytes to receive
 			 assert(recvCounter <= totalMessageSize);
 
 			 if (GetTimeBase() - startTime >= 1600) {
-				 printf("node %d: %llu bytes left\n", g_proc_id, recvCounter);
+				 //printf("node %d: %llu bytes left\n", g_proc_id, recvCounter);
 				 startTime = GetTimeBase();
 			 }
 		 }
-		 printf("node %d: All data received\n", g_proc_id);
+		 //printf("node %d: All data received\n", g_proc_id);
 
 		 // Wait for all data sent
 		while (true) {

diff --git a/bgq/bgq_operator.inc.c b/bgq/bgq_operator.inc.c
@@ -12,7 +12,7 @@
 
 #include <stdbool.h>
 
-
+#if 0
 #define OPERATOR_OUTPLACENAME bgq_operator_outplace
 #define OPERATOR_INPLACENAME bgq_operator_inplace
 #define OPERATOR_ARGFIELDS 1
@@ -182,3 +182,4 @@ void bgq_operator_inplace(bgq_weylfield_controlblock *targetfield) {
 
 #undef OPERATOR_INCLUDED
 #undef OPERATOR_NAME
+#endif