Add pci function ID to CudaDevice

breiters · Nov 15, 2024 · 7ee3196 · 7ee3196
1 parent f1c7420
commit 7ee3196
Show file tree

Hide file tree

Showing 7 changed files with 15 additions and 2 deletions.
diff --git a/doc/lua-doxygen.md b/doc/lua-doxygen.md
@@ -3723,6 +3723,10 @@ The option 'n' takes an argument, specified by the ':'. If found the option argu
       <TD>\a pciDom</TD>
       <TD>PCI domain identifier of the device</TD>
     </TR>
+    <TR>
+      <TD>\a pciFunc</TD>
+      <TD>PCI function identifier of the device</TD>
+    </TR>
     <TR>
       <TD>\a maxBlockRegs</TD>
       <TD>Maximum number of 32-bit registers available to a thread block</TD>

diff --git a/src/applications/likwid-topology.lua b/src/applications/likwid-topology.lua
@@ -329,6 +329,7 @@ if likwid.nvSupported() then
                 table.insert(output_csv, string.format("PCI bus:\t\t0x%x", gpu["pciBus"]))
                 table.insert(output_csv, string.format("PCI domain:\t\t0x%x", gpu["pciDom"]))
                 table.insert(output_csv, string.format("PCI device:\t\t0x%x", gpu["pciDev"]))
+                table.insert(output_csv, string.format("PCI function:\t\t0x%x", gpu["pciFunc"]))
             end
             table.insert(output_csv, likwid.hline)
         end

diff --git a/src/includes/likwid.h b/src/includes/likwid.h
@@ -2195,6 +2195,7 @@ typedef struct {
   int pciDev; /*!< \brief PCI device (also known as slot) identifier of the
                  device */
   int pciDom; /*!< \brief PCI domain identifier of the device */
+  int pciFunc; /*!< \brief PCI function identifier of the device */
   int maxBlockRegs;  /*!< \brief Maximum number of 32-bit registers available to
                         a thread block */
   int numMultiProcs; /*!< \brief Number of multiprocessors on the device */

diff --git a/src/likwid_device.c b/src/likwid_device.c
@@ -551,10 +551,11 @@ int likwid_device_get_available(LikwidDeviceType type, char ***id_list)
         {
 #ifdef LIKWID_WITH_NVMON
             case DEVICE_TYPE_NVIDIA_GPU:
-                snprintf(id_str, sizeof(id_str), "GN:%08x:%02x:%02x.0",
+                snprintf(id_str, sizeof(id_str), "GN:%08x:%02x:%02x.%01x",
                         cuda_topo->devices[i].pciDom,
                         cuda_topo->devices[i].pciBus,
-                        cuda_topo->devices[i].pciDev);
+                        cuda_topo->devices[i].pciDev,
+                        cuda_topo->devices[i].pciFunc);
                 break;
 #endif
 #ifdef LIKWID_WITH_ROCMON

diff --git a/src/luawid.c b/src/luawid.c
@@ -2625,6 +2625,9 @@ static int lua_likwid_getCudaTopology(lua_State *L) {
     lua_pushstring(L, "pciDom");
     lua_pushinteger(L, (lua_Integer)(gpu->pciDom));
     lua_settable(L, -3);
+    lua_pushstring(L, "pciFunc");
+    lua_pushinteger(L, (lua_Integer)(gpu->pciFunc));
+    lua_settable(L, -3);
     lua_pushstring(L, "maxBlockRegs");
     lua_pushinteger(L, (lua_Integer)(gpu->maxBlockRegs));
     lua_settable(L, -3);

diff --git a/src/topology_cuda.c b/src/topology_cuda.c
@@ -304,6 +304,8 @@ topology_cuda_init()
             CU_CALL((*cuDeviceGetAttributeTopoPtr)(&cudaTopology.devices[i].pciBus, CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, dev), ret = -ENOMEM; goto topology_gpu_init_error;);
             CU_CALL((*cuDeviceGetAttributeTopoPtr)(&cudaTopology.devices[i].pciDev, CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, dev), ret = -ENOMEM; goto topology_gpu_init_error;);
             CU_CALL((*cuDeviceGetAttributeTopoPtr)(&cudaTopology.devices[i].pciDom, CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, dev), ret = -ENOMEM; goto topology_gpu_init_error;);
+            // TODO: Get PCI function through nvmlDeviceGetPciInfo_v3, nvmlPciInfo_t->function
+            cudaTopology.devices[i].pciFunc = 0;
             CU_CALL((*cuDeviceGetAttributeTopoPtr)(&cudaTopology.devices[i].maxBlockRegs, CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK, dev), ret = -ENOMEM; goto topology_gpu_init_error;);
             CU_CALL((*cuDeviceGetAttributeTopoPtr)(&cudaTopology.devices[i].numMultiProcs, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, dev), ret = -ENOMEM; goto topology_gpu_init_error;);
             CU_CALL((*cuDeviceGetAttributeTopoPtr)(&cudaTopology.devices[i].maxThreadPerMultiProc, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR, dev), ret = -ENOMEM; goto topology_gpu_init_error;);

diff --git a/test/test-topology-gpu-rocm.c b/test/test-topology-gpu-rocm.c
@@ -45,6 +45,7 @@ int main()
         printf("pciBus: %d\n", device->pciBus);
         printf("pciDev: %d\n", device->pciDev);
         printf("pciDom: %d\n", device->pciDom);
+        printf("pciFunc: %d\n", device->pciFunc);
         printf("numMultiProcs: %d\n", device->numMultiProcs);
         printf("maxThreadPerMultiProc: %d\n", device->maxThreadPerMultiProc);
         printf("memBusWidth: %d\n", device->memBusWidth);