Merge branch 'assign_tasks_to_gpu' into 'master_pre_stage'

Assign tasks to gpu

See merge request !55

Makefile.am

@@ -138,7 +138,7 @@ libelpa@SUFFIX@_private_la_SOURCES += \
 endif
 
 if WITH_GPU_VERSION
-  libelpa@SUFFIX@_private_la_SOURCES += src/GPU/cudaFunctions.cu src/GPU/cuUtils.cu src/elpa2/GPU/ev_tridi_band_gpu_real.cu src/elpa2/GPU/ev_tridi_band_gpu_complex.cu
+  libelpa@SUFFIX@_private_la_SOURCES += src/GPU/elpa_index_gpu.cu src/GPU/cudaFunctions.cu src/GPU/cuUtils.cu src/elpa2/GPU/ev_tridi_band_gpu_real.cu src/elpa2/GPU/ev_tridi_band_gpu_complex.cu
 endif
 
 if !WITH_MPI

generate_automake_test_programs.py

@@ -24,6 +24,10 @@ gpu_flag = {
     0: "-DTEST_GPU=0",
     1: "-DTEST_GPU=1",
 }
+gpu_id_flag = {
+    0: "-DTEST_GPU_SET_ID=0",
+    1: "-DTEST_GPU_SET_ID=1",
+}
 
 matrix_flag = {
     "random":   "-DTEST_MATRIX_RANDOM",
@@ -57,9 +61,10 @@ split_comm_flag = {
     "by_elpa": ""
 }
 
-for lang, m, g, q, t, p, d, s, lay, spl in product(sorted(language_flag.keys()),
+for lang, m, g, gid, q, t, p, d, s, lay, spl in product(sorted(language_flag.keys()),
                                                    sorted(matrix_flag.keys()),
                                                    sorted(gpu_flag.keys()),
+                                                   sorted(gpu_id_flag.keys()),
                                                    sorted(qr_flag.keys()),
                                                    sorted(test_type_flag.keys()),
                                                    sorted(prec_flag.keys()),
@@ -68,6 +73,9 @@ for lang, m, g, q, t, p, d, s, lay, spl in product(sorted(language_flag.keys()),
                                                    sorted(layout_flag.keys()),
                                                    sorted(split_comm_flag.keys())):
 
+    if gid == 1 and (g == 0 ):
+        continue
+
     if lang == "C" and (m == "analytic" or m == "toeplitz" or m == "frank" or lay == "all_layouts"):
         continue
 
@@ -178,11 +186,12 @@ for lang, m, g, q, t, p, d, s, lay, spl in product(sorted(language_flag.keys()),
                 raise Exception("Oh no!")
             endifs += 1
 
-        name = "validate{langsuffix}_{d}_{p}_{t}_{s}{kernelsuffix}_{gpusuffix}{qrsuffix}{m}{layoutsuffix}{spl}".format(
+        name = "validate{langsuffix}_{d}_{p}_{t}_{s}{kernelsuffix}_{gpusuffix}{gpuidsuffix}{qrsuffix}{m}{layoutsuffix}{spl}".format(
             langsuffix=language_flag[lang],
             d=d, p=p, t=t, s=s,
             kernelsuffix="" if kernel == "nokernel" else "_" + kernel,
             gpusuffix="gpu_" if g else "",
+            gpuidsuffix="set_gpu_id_" if gid else "",
             qrsuffix="qr_" if q else "",
             m=m,
             layoutsuffix="_all_layouts" if lay == "all_layouts" else "",
@@ -227,6 +236,7 @@ for lang, m, g, q, t, p, d, s, lay, spl in product(sorted(language_flag.keys()),
             test_type_flag[t],
             solver_flag[s],
             gpu_flag[g],
+            gpu_id_flag[gid],
             qr_flag[q],
             matrix_flag[m]] + extra_flags))
 

src/GPU/check_for_gpu.F90

@@ -62,7 +62,7 @@ module mod_check_for_gpu
       integer(kind=ik), intent(out) :: numberOfDevices
       integer(kind=ik)              :: deviceNumber, mpierr, maxNumberOfDevices
       logical                       :: gpuAvailable
-      integer(kind=ik)              :: error, mpi_comm_all
+      integer(kind=ik)              :: error, mpi_comm_all, use_gpu_id, min_use_gpu_id
       !character(len=1024)           :: envname
 
       if (.not.(present(wantDebug))) then
@@ -83,42 +83,24 @@ module mod_check_for_gpu
         stop
       endif
 
-      if (cublasHandle .ne. -1) then
-        gpuAvailable = .true.
-        numberOfDevices = -1
-        if (myid == 0 .and. wantDebugMessage) then
-          print *, "Skipping GPU init, should have already been initialized "
-        endif
-        return
-      else
-        if (myid == 0 .and. wantDebugMessage) then
-          print *, "Initializing the GPU devices"
+      if (obj%is_set("use_gpu_id") == 1) then
+        call obj%get("use_gpu_id", use_gpu_id, error)
+        if (use_gpu_id == -99) then
+          print *,"Problem you did not set which gpu id this task should use"
         endif
-      endif
-
-      ! call getenv("CUDA_PROXY_PIPE_DIRECTORY", envname)
-      success = cuda_getdevicecount(numberOfDevices)
-
-      if (.not.(success)) then
-        print *,"error in cuda_getdevicecount"
-        stop 1
-      endif
-
-      ! make sure that all nodes have the same number of GPU's, otherwise
-      ! we run into loadbalancing trouble
+ 
+        ! check whether gpu ud has been set for each proces
 #ifdef WITH_MPI
-      call mpi_allreduce(numberOfDevices, maxNumberOfDevices, 1, MPI_INTEGER, MPI_MAX, mpi_comm_all, mpierr)
+        call mpi_allreduce(use_gpu_id, min_use_gpu_id, 1, MPI_INTEGER, MPI_MAX, mpi_comm_all, mpierr)
 
-      if (maxNumberOfDevices .ne. numberOfDevices) then
-        print *,"Different number of GPU devices on MPI tasks!"
-        print *,"GPUs will NOT be used!"
-        gpuAvailable = .false.
-        return
-      endif
+        if (min_use_gpu_id .lt. 0) then
+          print *,"Not all tasks have set which GPU id should be used"
+          print *,"GPUs will NOT be used!"
+          gpuAvailable = .false.
+          return
+        endif
 #endif
-      if (numberOfDevices .ne. 0) then
         gpuAvailable = .true.
-        ! Usage of GPU is possible since devices have been detected
 
         if (myid==0) then
           if (wantDebugMessage) then
@@ -127,8 +109,7 @@ module mod_check_for_gpu
           endif
         endif
 
-        deviceNumber = mod(myid, numberOfDevices)
-        success = cuda_setdevice(deviceNumber)
+        success = cuda_setdevice(use_gpu_id)
 
         if (.not.(success)) then
           print *,"Cannot set CudaDevice"
@@ -137,14 +118,77 @@ module mod_check_for_gpu
         if (wantDebugMessage) then
           print '(3(a,i0))', 'MPI rank ', myid, ' uses GPU #', deviceNumber
         endif
-        
+          
         success = cublas_create(cublasHandle)
         if (.not.(success)) then
           print *,"Cannot create cublas handle"
           stop 1
         endif
-        
-      endif
+      else
+
+        if (cublasHandle .ne. -1) then
+          gpuAvailable = .true.
+          numberOfDevices = -1
+          if (myid == 0 .and. wantDebugMessage) then
+            print *, "Skipping GPU init, should have already been initialized "
+          endif
+          return
+        else
+          if (myid == 0 .and. wantDebugMessage) then
+            print *, "Initializing the GPU devices"
+          endif
+        endif
+
+        ! call getenv("CUDA_PROXY_PIPE_DIRECTORY", envname)
+        success = cuda_getdevicecount(numberOfDevices)
+
+        if (.not.(success)) then
+          print *,"error in cuda_getdevicecount"
+          stop 1
+        endif
+
+        ! make sure that all nodes have the same number of GPU's, otherwise
+        ! we run into loadbalancing trouble
+#ifdef WITH_MPI
+        call mpi_allreduce(numberOfDevices, maxNumberOfDevices, 1, MPI_INTEGER, MPI_MAX, mpi_comm_all, mpierr)
+
+        if (maxNumberOfDevices .ne. numberOfDevices) then
+          print *,"Different number of GPU devices on MPI tasks!"
+          print *,"GPUs will NOT be used!"
+          gpuAvailable = .false.
+          return
+        endif
+#endif
+        if (numberOfDevices .ne. 0) then
+          gpuAvailable = .true.
+          ! Usage of GPU is possible since devices have been detected
+
+          if (myid==0) then
+            if (wantDebugMessage) then
+              print *
+              print '(3(a,i0))','Found ', numberOfDevices, ' GPUs'
+            endif
+          endif
+
+          deviceNumber = mod(myid, numberOfDevices)
+          success = cuda_setdevice(deviceNumber)
+
+          if (.not.(success)) then
+            print *,"Cannot set CudaDevice"
+            stop 1
+          endif
+          if (wantDebugMessage) then
+            print '(3(a,i0))', 'MPI rank ', myid, ' uses GPU #', deviceNumber
+          endif
+          
+          success = cublas_create(cublasHandle)
+          if (.not.(success)) then
+            print *,"Cannot create cublas handle"
+            stop 1
+          endif
+          
+        endif
 
+      endif  
     end function
 end module

src/GPU/elpa_index_gpu.cu

+extern "C" {
+  int gpu_count() {
+    int count;
+    cudaError_t cuerr = cudaGetDeviceCount(&count);
+    if (cuerr != cudaSuccess) {
+      count = -1000;
+    }
+    return count;
+  }
+}

src/elpa.F90

@@ -143,6 +143,16 @@
 !>     print *,"Could not setup ELPA object"
 !>   endif
 !>
+!>   ! settings for GPU
+!>   call elpaInstance%set("gpu", 1, success) ! 1=on, 2=off
+!>   ! in case of GPU usage you have the choice whether ELPA
+!>   ! should automatically assign each MPI task to a certain GPU
+!>   ! (this is default) or whether you want to set this assignment
+!>   ! for _each_ task yourself
+!>   ! set assignment your self (only using one task here and assigning it 
+!>   ! to GPU id 1)
+!>   if (my_rank .eq. 0) call elpaInstance%set("use_gpu_id", 1, success)
+!>
 !>   ! if desired, set tunable run-time options
 !>   ! here we want to use the 2-stage solver
 !>   call elpaInstance%set("solver", ELPA_SOLVER_2STAGE, success)
@@ -206,6 +216,16 @@
 !>   /* here we want to use the 2-stage solver */
 !>   elpa_set(handle, "solver", ELPA_SOLVER_2STAGE, &error);
 !>
+!>   /* settings for GPU */
+!>   elpa_set(handle, "gpu", 1, &error);  /* 1=on, 2=off */
+!>   /* in case of GPU usage you have the choice whether ELPA
+!>      should automatically assign each MPI task to a certain GPU
+!>      (this is default) or whether you want to set this assignment
+!>      for _each_ task yourself
+!>      set assignment your self (only using one task here and assigning it 
+!>      to GPU id 1) */
+!>   if (my_rank == 0) elpa_set(handle, "use_gpu_id", 1, &error);
+!>
 !>   elpa_set(handle,"real_kernel", ELPA_2STAGE_REAL_AVX_BLOCK2, &error);
 !>  \endcode
 !>   ... set and get all other options that are desired

src/elpa_index.c

@@ -112,6 +112,14 @@ static int min_tile_size_cardinality(elpa_index_t index);
 static int min_tile_size_enumerate(elpa_index_t index, int i);
 static int min_tile_size_is_valid(elpa_index_t index, int n, int new_value);
 
+#ifdef WITH_GPU_VERSION
+int gpu_count();
+#endif
+
+static int use_gpu_id_cardinality(elpa_index_t index);
+static int use_gpu_id_enumerate(elpa_index_t index, int i);
+static int use_gpu_id_is_valid(elpa_index_t index, int n, int new_value);
+
 static int valid_with_gpu(elpa_index_t index, int n, int new_value);
 static int valid_with_gpu_elpa1(elpa_index_t index, int n, int new_value);
 static int valid_with_gpu_elpa2(elpa_index_t index, int n, int new_value);
@@ -236,6 +244,8 @@ static const elpa_index_int_entry_t int_entries[] = {
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa2, NULL, PRINT_YES),
         INT_ENTRY("gpu_trans_ev_band_to_full", "Use GPU acceleration for ELPA2 trans_ev_band_to_full", 1, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY, \
                         cardinality_bool, enumerate_identity, valid_with_gpu_elpa2, NULL, PRINT_YES),
+	INT_ENTRY("use_gpu_id", "Calling MPI task will use this gpu id", -99, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY, \
+		  use_gpu_id_cardinality, use_gpu_id_enumerate, use_gpu_id_is_valid, NULL, PRINT_YES), 
         INT_ENTRY("real_kernel", "Real kernel to use if 'solver' is set to ELPA_SOLVER_2STAGE", ELPA_2STAGE_REAL_DEFAULT, ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_REAL, \
                         number_of_real_kernels, real_kernel_enumerate, real_kernel_is_valid, real_kernel_name, PRINT_YES),
         INT_ENTRY("complex_kernel", "Complex kernel to use if 'solver' is set to ELPA_SOLVER_2STAGE", ELPA_2STAGE_COMPLEX_DEFAULT, ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_COMPLEX, \
@@ -1093,6 +1103,40 @@ static int max_stored_rows_is_valid(elpa_index_t index, int n, int new_value) {
         }
 }
 
+static int use_gpu_id_cardinality(elpa_index_t index) {
+#ifdef WITH_GPU_VERSION
+	int count;
+	count = gpu_count();
+        if (count == -1000) {
+          fprintf(stderr, "Querrying GPUs failed! Set GPU count = 0\n");
+	return 0;
+        }
+	return count;
+#else
+	return 0;
+#endif
+}
+
+static int use_gpu_id_enumerate(elpa_index_t index, int i) {
+        fprintf(stderr, "use_gpu_id_is_enumerate should never be called. please report this bug\n");
+        return i;
+}
+
+static int use_gpu_id_is_valid(elpa_index_t index, int n, int new_value) {
+#ifdef WITH_GPU_VERSION
+	int count;
+	count = gpu_count();
+        if (count == -1000) {
+          fprintf(stderr, "Querrying GPUs failed! Return with error\n");
+	  return 0 == 1 ;
+	} else {
+          return (0 <= new_value) && (new_value <= count);
+	}
+#else
+	return 0 == 0;
+#endif
+
+}
 
 // TODO: this shoudl definitely be improved (too many options to test in autotuning)
 static const int TILE_SIZE_STEP = 128;

test/Fortran/test.F90

@@ -199,7 +199,9 @@ program test
 #ifdef WITH_OPENMP_TRADITIONAL
    TEST_INT_TYPE      :: max_threads, threads_caller
 #endif
-
+#ifdef TEST_GPU_SET_ID
+   TEST_INT_TYPE      :: gpuID
+#endif
 #ifdef SPLIT_COMM_MYSELF
    TEST_INT_MPI_TYPE  :: mpi_comm_rows, mpi_comm_cols, mpi_string_length, mpierr2
    character(len=MPI_MAX_ERROR_STRING) :: mpierr_string
@@ -641,6 +643,16 @@ program test
    call e%set("gpu", TEST_GPU, error_elpa)
    assert_elpa_ok(error_elpa)
 
+#ifdef TEST_GPU_SET_ID
+   ! simple test
+   ! Can (and should) fail often
+   gpuID = mod(myid,2)
+   print *,"Task",myid,"wants to use GPU",gpuID
+   call e%set("use_gpu_id", int(gpuID,kind=c_int), error_elpa)
+   assert_elpa_ok(error_elpa)
+#endif
+
+
 #if TEST_QR_DECOMPOSITION == 1
    call e%set("qr", 1_ik, error_elpa)
    assert_elpa_ok(error_elpa)