Gcn arch name (#886)

We use CMake to determine if we're compiling against a version of ROCm that supports gcnArchName and handles architecture checking appropriately. It includes a few helper functions as drop ins for the functionality we used gcnArch for before; sometimes to enable flags, and sometimes to set frequencies.

src/graph/topo.cc

@@ -369,7 +369,14 @@ ncclResult_t ncclTopoAddNic(struct ncclXmlNode* xmlNic, struct ncclTopoSystem* s
 
 ncclResult_t ncclTopoAddGpu(struct ncclXmlNode* xmlGpu, struct ncclTopoSystem* system, struct ncclTopoNode* gpu) {
   NCCLCHECK(xmlGetAttrInt(xmlGpu, "sm", &gpu->gpu.cudaCompCap));
-  NCCLCHECK(xmlGetAttrInt(xmlGpu, "gcn", &gpu->gpu.gcn));
+  NCCLCHECK(xmlGetAttr(xmlGpu, "gcn", &gpu->gpu.gcn));
+  if (strcmp(gpu->gpu.gcn, "906") == 0) {
+    gpu->gpu.gcn = "gfx906";
+  } else if (strcmp(gpu->gpu.gcn, "908") == 0) {
+    gpu->gpu.gcn = "gfx908";
+  } else if (strcmp(gpu->gpu.gcn, "910") == 0) {
+    gpu->gpu.gcn = "gfx90a";
+  }
   rcclHipDeviceArch_t arch;
   NCCLCHECK(xmlGetAttrInt(xmlGpu, "arch", &arch.value));
   memcpy(&gpu->gpu.arch, &arch.arch, sizeof(hipDeviceArch_t));

src/graph/topo.h

@@ -10,6 +10,8 @@
 
 #include "graph.h"
 #include "core.h"
+#include "archinfo.h"
+#include <string.h>
 
 #define LOC_BW 5000.0
 #define SM60_NVLINK_BW 18.0
@@ -123,7 +125,7 @@ struct ncclTopoNode {
       int rank;
       int cudaCompCap;
       int gdrSupport;
-      int gcn;
+      const char* gcn;
       hipDeviceArch_t arch;
     }gpu;
     struct {
@@ -224,17 +226,13 @@ static ncclResult_t ncclTopoDevToRank(struct ncclTopoSystem* system, int dev, in
 }
 
 // Returns XGMI speed in GB/s
-static float ncclTopoXGMISpeed(int gcn) {
-  switch (gcn) {
-    case 910:
-      return MI200_XGMI_WIDTH;
-    case 940:
-    case 941:
-    case 942:
-      return GFX94X_XGMI_WIDTH;
-    default:
-      return VEGA_XGMI_WIDTH;
-  }
+static float ncclTopoXGMISpeed(const char* gcn) {
+  if (IsArchMatch(gcn, "gfx90a"))
+    return MI200_XGMI_WIDTH;
+  else if (IsArchMatch(gcn, "gfx94"))
+    return GFX94X_XGMI_WIDTH;
+  else
+    return VEGA_XGMI_WIDTH;
 }
 
 #if ENABLE_COLLTRACE

src/graph/tuning.cc

@@ -455,7 +455,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
 #if defined(ENABLE_LL128)
       // Enable LL128 by default only on gfx90a with available tuning table
       pEnable = (graphs[a]->typeInter <= PATH_PXB) && graphs[a]->typeIntra <= PATH_NVL &&
-        (comm->topo->nodes[GPU].nodes[0].gpu.gcn == 910 && comm->topo->ll128Enabled) ? 1 : 0;
+        (IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx90a") && comm->topo->ll128Enabled) ? 1 : 0;
 #else
       pEnable = 0;
 #endif

src/graph/xml.cc

@@ -14,6 +14,7 @@
 #include "nvmlwrap.h"
 #include "xml.h"
 #include "rocm_smi_wrap.h"
+#include "archinfo.h"
 
 /*******************/
 /* XML File Parser */

src/graph/xml.h

@@ -12,6 +12,7 @@
 #include "debug.h"
 #include "checks.h"
 #include <stdlib.h>
+#include "archinfo.h"
 
 // A few constraints to make the implementation easy
 #define MAX_STR_LEN 255