Merge pull request #740 from CEED/natalie/device-id

Update device ID selection for HIP/CUDA/MAGMA backends

README.rst

@@ -265,6 +265,11 @@ Currently, each MAGMA library installation is only built for either CUDA or HIP.
 set of libCEED backends (``/gpu/cuda/magma/*`` or ``/gpu/hip/magma/*``) will automatically be built
 for the version of the MAGMA library found in ``MAGMA_DIR``.
 
+Users can specify a device for all CUDA, HIP, and MAGMA backends through adding `:device_id=#` 
+after the resource name.  For example:
+
+ - `/gpu/cuda/gen:device_id=1`
+
 The ``/*/occa`` backends rely upon the `OCCA <http://github.com/libocca/occa>`_ package to provide
 cross platform performance. To enable the OCCA backend, the environment variable ``OCCA_DIR`` must point
 to the top-level OCCA directory, with the OCCA library located in the ``${OCCA_DIR}/lib`` (By default,

backends/cuda-gen/ceed-cuda-gen.c

@@ -32,15 +32,15 @@ static int CeedInit_Cuda_gen(const char *resource, Ceed ceed) {
                      "Cuda backend cannot use resource: %s", resource);
   // LCOV_EXCL_STOP
 
-  Ceed ceedshared;
-  CeedInit("/gpu/cuda/shared", &ceedshared);
-  ierr = CeedSetDelegate(ceed, ceedshared); CeedChkBackend(ierr);
-
-  Ceed_Cuda_gen *data;
+  Ceed_Cuda *data;
   ierr = CeedCalloc(1, &data); CeedChkBackend(ierr);
   ierr = CeedSetData(ceed, data); CeedChkBackend(ierr);
   ierr = CeedCudaInit(ceed, resource, nrc); CeedChkBackend(ierr);
 
+  Ceed ceedshared;
+  CeedInit("/gpu/cuda/shared", &ceedshared);
+  ierr = CeedSetDelegate(ceed, ceedshared); CeedChkBackend(ierr);
+
   const char fallbackresource[] = "/gpu/cuda/ref";
   ierr = CeedSetOperatorFallbackResource(ceed, fallbackresource);
   CeedChkBackend(ierr);

backends/cuda-gen/ceed-cuda-gen.h

@@ -44,10 +44,6 @@ typedef struct {
   void *d_c;
 } CeedQFunction_Cuda_gen;
 
-typedef struct {
-  Ceed_Cuda base;
-} Ceed_Cuda_gen;
-
 CEED_INTERN int CeedQFunctionCreate_Cuda_gen(CeedQFunction qf);
 
 CEED_INTERN int CeedOperatorCreate_Cuda_gen(CeedOperator op);

backends/cuda-shared/ceed-cuda-shared-basis.c

@@ -773,7 +773,7 @@ int CeedBasisApplyTensor_Cuda_shared(CeedBasis basis, const CeedInt nelem,
   int ierr;
   Ceed ceed;
   ierr = CeedBasisGetCeed(basis, &ceed); CeedChkBackend(ierr);
-  Ceed_Cuda_shared *ceed_Cuda;
+  Ceed_Cuda *ceed_Cuda;
   CeedGetData(ceed, &ceed_Cuda); CeedChkBackend(ierr);
   CeedBasis_Cuda_shared *data;
   CeedBasisGetData(basis, &data); CeedChkBackend(ierr);

backends/cuda-shared/ceed-cuda-shared.c

@@ -16,7 +16,6 @@
 
 #include <ceed/ceed.h>
 #include <ceed/backend.h>
-#include <stdbool.h>
 #include <string.h>
 #include "ceed-cuda-shared.h"
 #include "../cuda/ceed-cuda.h"
@@ -34,15 +33,15 @@ static int CeedInit_Cuda_shared(const char *resource, Ceed ceed) {
   // LCOV_EXCL_STOP
   ierr = CeedSetDeterministic(ceed, true); CeedChk(ierr);
 
-  Ceed ceedref;
-  CeedInit("/gpu/cuda/ref", &ceedref);
-  ierr = CeedSetDelegate(ceed, ceedref); CeedChk(ierr);
-
-  Ceed_Cuda_shared *data;
+  Ceed_Cuda *data;
   ierr = CeedCalloc(1, &data); CeedChk(ierr);
   ierr = CeedSetData(ceed, data); CeedChk(ierr);
   ierr = CeedCudaInit(ceed, resource, nrc); CeedChk(ierr);
 
+  Ceed ceedref;
+  CeedInit("/gpu/cuda/ref", &ceedref);
+  ierr = CeedSetDelegate(ceed, ceedref); CeedChk(ierr);
+
   ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateTensorH1",
                                 CeedBasisCreateTensorH1_Cuda_shared);
   CeedChk(ierr);

backends/cuda-shared/ceed-cuda-shared.h

@@ -35,10 +35,6 @@ typedef struct {
   CeedScalar *c_G;
 } CeedBasis_Cuda_shared;
 
-typedef struct {
-  Ceed_Cuda base;
-} Ceed_Cuda_shared;
-
 CEED_INTERN int CeedBasisCreateTensorH1_Cuda_shared(CeedInt dim, CeedInt P1d,
     CeedInt Q1d, const CeedScalar *interp1d, const CeedScalar *grad1d,
     const CeedScalar *qref1d, const CeedScalar *qweight1d, CeedBasis basis);

backends/cuda/ceed-cuda.c

@@ -21,7 +21,6 @@
 #include <cuda_runtime.h>
 #include <nvrtc.h>
 #include <stdarg.h>
-#include <stdbool.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -150,22 +149,22 @@ static int CeedGetPreferredMemType_Cuda(CeedMemType *type) {
 //------------------------------------------------------------------------------
 int CeedCudaInit(Ceed ceed, const char *resource, int nrc) {
   int ierr;
-  const int rlen = strlen(resource);
-  const bool slash = (rlen>nrc) ? (resource[nrc] == '/') : false;
-  const int deviceID = (slash && rlen > nrc + 1) ? atoi(&resource[nrc + 1]) : -1;
+  const char *device_spec = strstr(resource, ":device_id=");
+  const int deviceID = (device_spec) ? atoi(device_spec+11) : -1;
 
   int currentDeviceID;
   ierr = cudaGetDevice(&currentDeviceID); CeedChk_Cu(ceed,ierr);
   if (deviceID >= 0 && currentDeviceID != deviceID) {
     ierr = cudaSetDevice(deviceID); CeedChk_Cu(ceed,ierr);
+    currentDeviceID = deviceID;
   }
-
   struct cudaDeviceProp deviceProp;
-  ierr = cudaGetDeviceProperties(&deviceProp, deviceID); CeedChk_Cu(ceed,ierr);
+  ierr = cudaGetDeviceProperties(&deviceProp, currentDeviceID);
+  CeedChk_Cu(ceed,ierr);
 
   Ceed_Cuda *data;
   ierr = CeedGetData(ceed, &data); CeedChkBackend(ierr);
-  data->deviceId = deviceID;
+  data->deviceId = currentDeviceID;
   data->optblocksize = deviceProp.maxThreadsPerBlock;
   return CEED_ERROR_SUCCESS;
 }
@@ -210,6 +209,7 @@ static int CeedInit_Cuda(const char *resource, Ceed ceed) {
     return CeedError(ceed, CEED_ERROR_BACKEND,
                      "Cuda backend cannot use resource: %s", resource);
   // LCOV_EXCL_STOP
+  ierr = CeedSetDeterministic(ceed, true); CeedChk(ierr);
 
   Ceed_Cuda *data;
   ierr = CeedCalloc(1, &data); CeedChkBackend(ierr);

backends/hip-gen/ceed-hip-gen.c

@@ -32,15 +32,15 @@ static int CeedInit_Hip_gen(const char *resource, Ceed ceed) {
                      "Hip backend cannot use resource: %s", resource);
   // LCOV_EXCL_STOP
 
-  Ceed ceedshared;
-  CeedInit("/gpu/hip/shared", &ceedshared);
-  ierr = CeedSetDelegate(ceed, ceedshared); CeedChkBackend(ierr);
-
-  Ceed_Hip_gen *data;
+  Ceed_Hip *data;
   ierr = CeedCalloc(1, &data); CeedChkBackend(ierr);
   ierr = CeedSetData(ceed, data); CeedChkBackend(ierr);
   ierr = CeedHipInit(ceed, resource, nrc); CeedChkBackend(ierr);
 
+  Ceed ceedshared;
+  CeedInit("/gpu/hip/shared", &ceedshared);
+  ierr = CeedSetDelegate(ceed, ceedshared); CeedChkBackend(ierr);
+
   const char fallbackresource[] = "/gpu/hip/ref";
   ierr = CeedSetOperatorFallbackResource(ceed, fallbackresource);
   CeedChkBackend(ierr);

backends/hip-gen/ceed-hip-gen.h

@@ -44,10 +44,6 @@ typedef struct {
   void *d_c;
 } CeedQFunction_Hip_gen;
 
-typedef struct {
-  Ceed_Hip base;
-} Ceed_Hip_gen;
-
 CEED_INTERN int CeedQFunctionCreate_Hip_gen(CeedQFunction qf);
 
 CEED_INTERN int CeedOperatorCreate_Hip_gen(CeedOperator op);

backends/hip-shared/ceed-hip-shared-basis.c

@@ -859,7 +859,7 @@ int CeedBasisApplyTensor_Hip_shared(CeedBasis basis, const CeedInt nelem,
   int ierr;
   Ceed ceed;
   ierr = CeedBasisGetCeed(basis, &ceed); CeedChkBackend(ierr);
-  Ceed_Hip_shared *ceed_Hip;
+  Ceed_Hip *ceed_Hip;
   CeedGetData(ceed, &ceed_Hip); CeedChkBackend(ierr);
   CeedBasis_Hip_shared *data;
   CeedBasisGetData(basis, &data); CeedChkBackend(ierr);

backends/hip-shared/ceed-hip-shared.c

@@ -34,15 +34,15 @@ static int CeedInit_Hip_shared(const char *resource, Ceed ceed) {
   // LCOV_EXCL_STOP
   ierr = CeedSetDeterministic(ceed, true); CeedChkBackend(ierr);
 
-  Ceed ceedref;
-  CeedInit("/gpu/hip/ref", &ceedref);
-  ierr = CeedSetDelegate(ceed, ceedref); CeedChkBackend(ierr);
-
-  Ceed_Hip_shared *data;
+  Ceed_Hip *data;
   ierr = CeedCalloc(1, &data); CeedChkBackend(ierr);
   ierr = CeedSetData(ceed, data); CeedChkBackend(ierr);
   ierr = CeedHipInit(ceed, resource, nrc); CeedChkBackend(ierr);
 
+  Ceed ceedref;
+  CeedInit("/gpu/hip/ref", &ceedref);
+  ierr = CeedSetDelegate(ceed, ceedref); CeedChkBackend(ierr);
+
   ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "BasisCreateTensorH1",
                                 CeedBasisCreateTensorH1_Hip_shared);
   CeedChkBackend(ierr);

backends/hip-shared/ceed-hip-shared.h

@@ -34,10 +34,6 @@ typedef struct {
   CeedScalar *d_qweight1d;
 } CeedBasis_Hip_shared;
 
-typedef struct {
-  Ceed_Hip base;
-} Ceed_Hip_shared;
-
 CEED_INTERN int CeedBasisCreateTensorH1_Hip_shared(CeedInt dim, CeedInt P1d,
     CeedInt Q1d, const CeedScalar *interp1d, const CeedScalar *grad1d,
     const CeedScalar *qref1d, const CeedScalar *qweight1d, CeedBasis basis);

backends/hip/ceed-hip.c

@@ -16,8 +16,8 @@
 
 #include <ceed/ceed.h>
 #include <ceed/backend.h>
-#include <stdarg.h>
 #include <string.h>
+#include <stdlib.h>
 #include "ceed-hip.h"
 
 //------------------------------------------------------------------------------
@@ -33,22 +33,23 @@ static int CeedGetPreferredMemType_Hip(CeedMemType *type) {
 //------------------------------------------------------------------------------
 int CeedHipInit(Ceed ceed, const char *resource, int nrc) {
   int ierr;
-  const int rlen = strlen(resource);
-  const bool slash = (rlen>nrc) ? (resource[nrc] == '/') : false;
-  const int deviceID = (slash && rlen > nrc + 1) ? atoi(&resource[nrc + 1]) : -1;
+  const char *device_spec = strstr(resource, ":device_id=");
+  const int deviceID = (device_spec) ? atoi(device_spec+11) : -1;
 
   int currentDeviceID;
   ierr = hipGetDevice(&currentDeviceID); CeedChk_Hip(ceed,ierr);
   if (deviceID >= 0 && currentDeviceID != deviceID) {
     ierr = hipSetDevice(deviceID); CeedChk_Hip(ceed,ierr);
+    currentDeviceID = deviceID;
   }
 
   struct hipDeviceProp_t deviceProp;
-  ierr = hipGetDeviceProperties(&deviceProp, deviceID); CeedChk_Hip(ceed,ierr);
+  ierr = hipGetDeviceProperties(&deviceProp, currentDeviceID);
+  CeedChk_Hip(ceed,ierr);
 
   Ceed_Hip *data;
   ierr = CeedGetData(ceed, &data); CeedChkBackend(ierr);
-  data->deviceId = deviceID;
+  data->deviceId = currentDeviceID;
   data->optblocksize = 256;
   return CEED_ERROR_SUCCESS;
 }
@@ -93,6 +94,7 @@ static int CeedInit_Hip(const char *resource, Ceed ceed) {
     return CeedError(ceed, CEED_ERROR_BACKEND,
                      "Hip backend cannot use resource: %s", resource);
   // LCOV_EXCL_STOP
+  ierr = CeedSetDeterministic(ceed, true); CeedChk(ierr);
 
   Ceed_Hip *data;
   ierr = CeedCalloc(1, &data); CeedChkBackend(ierr);

backends/magma/ceed-magma-det.c

@@ -16,18 +16,38 @@
 
 #include <ceed/ceed.h>
 #include <ceed/backend.h>
+#include <string.h>
+#include <stdlib.h>
 #include "ceed-magma.h"
 
 CEED_INTERN int CeedInit_Magma_Det(const char *resource, Ceed ceed) {
   int ierr;
-  if (strcmp(resource, "/gpu/cuda/magma/det")
-      && strcmp(resource, "/gpu/hip/magma/det"))
+  const int nrc = 18; // number of characters in resource
+  if (strncmp(resource, "/gpu/cuda/magma/det", nrc)
+      && strncmp(resource, "/gpu/hip/magma/det", nrc))
     // LCOV_EXCL_START
     return CeedError(ceed, CEED_ERROR_BACKEND,
                      "Magma backend cannot use resource: %s", resource);
   // LCOV_EXCL_STOP
   ierr = CeedSetDeterministic(ceed, true); CeedChkBackend(ierr);
 
+  Ceed_Magma *data;
+  ierr = CeedCalloc(sizeof(Ceed_Magma), &data); CeedChkBackend(ierr);
+  ierr = CeedSetData(ceed, data); CeedChkBackend(ierr);
+
+  // get/set device ID
+  const char *device_spec = strstr(resource, ":device_id=");
+  const int deviceID = (device_spec) ? atoi(device_spec+11) : -1;
+
+  int currentDeviceID;
+  magma_getdevice(&currentDeviceID);
+  if (deviceID >= 0 && currentDeviceID != deviceID) {
+    magma_setdevice(deviceID);
+    currentDeviceID = deviceID;
+  }
+  // create a queue that uses the null stream
+  data->device = currentDeviceID;
+
   // Create reference CEED that implementation will be dispatched
   //   through unless overridden
   Ceed ceedref;

backends/magma/ceed-magma-restriction.c

@@ -16,6 +16,7 @@
 
 #include <ceed/ceed.h>
 #include <ceed/backend.h>
+#include <string.h>
 #include "ceed-magma.h"
 
 static int CeedElemRestrictionApply_Magma(CeedElemRestriction r,

backends/magma/ceed-magma.c

@@ -17,6 +17,7 @@
 #include <ceed/ceed.h>
 #include <ceed/backend.h>
 #include <string.h>
+#include <stdlib.h>
 #include "ceed-magma.h"
 
 static int CeedDestroy_Magma(Ceed ceed) {
@@ -30,22 +31,14 @@ static int CeedDestroy_Magma(Ceed ceed) {
 
 static int CeedInit_Magma(const char *resource, Ceed ceed) {
   int ierr;
-  if (strcmp(resource, "/gpu/cuda/magma") && strcmp(resource, "/gpu/hip/magma"))
+  const int nrc = 14; // number of characters in resource
+  if (strncmp(resource, "/gpu/cuda/magma", nrc)
+      && strncmp(resource, "/gpu/hip/magma", nrc))
     // LCOV_EXCL_START
     return CeedError(ceed, CEED_ERROR_BACKEND,
                      "Magma backend cannot use resource: %s", resource);
   // LCOV_EXCL_STOP
 
-  // Create reference CEED that implementation will be dispatched
-  //   through unless overridden
-  Ceed ceedref;
-  #ifdef HAVE_HIP
-  CeedInit("/gpu/hip/ref", &ceedref);
-  #else
-  CeedInit("/gpu/cuda/ref", &ceedref);
-  #endif
-  ierr = CeedSetDelegate(ceed, ceedref); CeedChkBackend(ierr);
-
   ierr = magma_init();
   if (ierr)
     // LCOV_EXCL_START
@@ -64,14 +57,34 @@ static int CeedInit_Magma(const char *resource, Ceed ceed) {
   data->maxthreads[1] = 128;  // for 2D kernels
   data->maxthreads[2] =  64;  // for 3D kernels
 
+  // get/set device ID
+  const char *device_spec = strstr(resource, ":device_id=");
+  const int deviceID = (device_spec) ? atoi(device_spec+11) : -1;
+
+  int currentDeviceID;
+  magma_getdevice(&currentDeviceID);
+  if (deviceID >= 0 && currentDeviceID != deviceID) {
+    magma_setdevice(deviceID);
+    currentDeviceID = deviceID;
+  }
   // create a queue that uses the null stream
-  magma_getdevice( &(data->device) );
+  data->device = currentDeviceID;
   #ifdef HAVE_HIP
   magma_queue_create_from_hip(data->device, NULL, NULL, NULL, &(data->queue));
   #else
   magma_queue_create_from_cuda(data->device, NULL, NULL, NULL, &(data->queue));
   #endif
 
+  // Create reference CEED that implementation will be dispatched
+  //   through unless overridden
+  Ceed ceedref;
+  #ifdef HAVE_HIP
+  CeedInit("/gpu/hip/ref", &ceedref);
+  #else
+  CeedInit("/gpu/cuda/ref", &ceedref);
+  #endif
+  ierr = CeedSetDelegate(ceed, ceedref); CeedChkBackend(ierr);
+
   ierr = CeedSetBackendFunction(ceed, "Ceed", ceed, "ElemRestrictionCreate",
                                 CeedElemRestrictionCreate_Magma); CeedChkBackend(ierr);
   ierr = CeedSetBackendFunction(ceed, "Ceed", ceed,