gfx12 Disable ll protocol (#1268)

test/StandaloneTests.cpp

@@ -7,9 +7,10 @@
 #include <gtest/gtest.h>
 #include <rccl/rccl.h>
 
+#include "TestBed.hpp"
 #include "StandaloneUtils.hpp"
 
-namespace RcclUnitTesting 
+namespace RcclUnitTesting
 {
   /**
    * \brief Verify that each device is assigned to the right rank using ncclCommSplit API.
@@ -73,7 +74,7 @@ namespace RcclUnitTesting
     NCCLCHECK(ncclCommInitAll(comms.data(), numDevices, nullptr));
 
     // Split into new comms (all of the same color)
-    std::vector<ncclComm_t> subComms(numDevices);  
+    std::vector<ncclComm_t> subComms(numDevices);
     NCCLCHECK(ncclGroupStart());
     for (int localRank = 0; localRank < numDevices; localRank++)
       NCCLCHECK(ncclCommSplit(comms[localRank], 0, localRank, &subComms[localRank], NULL));
@@ -88,7 +89,7 @@ namespace RcclUnitTesting
       int subCommRank, subCommNRank;
       NCCLCHECK(ncclCommUserRank(subComms[i], &subCommRank));
       NCCLCHECK(ncclCommCount(subComms[i], &subCommNRank));
-          
+
       ASSERT_EQ(originalRank, subCommRank);
       ASSERT_EQ(originalNRank, subCommNRank);
     }
@@ -117,7 +118,7 @@ namespace RcclUnitTesting
     NCCLCHECK(ncclCommInitAll(comms.data(), numDevices, nullptr));
 
     // Split into new comms
-    int numReducedRanks = numDevices / 2; 
+    int numReducedRanks = numDevices / 2;
     std::vector<ncclComm_t> subComms(numDevices);
     NCCLCHECK(ncclGroupStart());
     for (int localRank = 0; localRank < numDevices; localRank++)
@@ -131,12 +132,12 @@ namespace RcclUnitTesting
       int originalRank, originalNRank;
       NCCLCHECK(ncclCommUserRank(comms[i], &originalRank));
       NCCLCHECK(ncclCommCount(comms[i], &originalNRank));
-        
+
       if (i < numReducedRanks) {
         int subCommRank, subCommNRank;
         NCCLCHECK(ncclCommUserRank(subComms[i], &subCommRank));
         NCCLCHECK(ncclCommCount(subComms[i], &subCommNRank));
-        
+
         ASSERT_EQ(originalRank, subCommRank);
         ASSERT_EQ(subCommNRank, numReducedRanks);
       } else {
@@ -150,12 +151,13 @@ namespace RcclUnitTesting
     for (auto& comm : comms)
       NCCLCHECK(ncclCommDestroy(comm));
   }
-  
+
   /**
    * \brief Verify there is no regression in timing for each protocol [LL, LL128, Simple]
    * ******************************************************************************************/
   TEST(Standalone, RegressionTiming)
   {
+    TestBed testBed;
     // timing
     using namespace std::chrono;
     using Clock = std::chrono::high_resolution_clock;
@@ -178,7 +180,12 @@ namespace RcclUnitTesting
     for (auto p : protocolList)
     {
       usElapsed = 0;
-      setenv("NCCL_PROTO", p, 1);
+      if(testBed.ev.isGfx12) {
+        setenv("NCCL_PROTO", "Simple", 1);
+      } else {
+        setenv("NCCL_PROTO", p, 1);
+      }
+
       NCCLCHECK(ncclCommInitAll(comms.data(), numRanks, nullptr));
 
       // Prepare CPU data arrays

test/common/EnvVars.cpp

@@ -15,7 +15,7 @@ namespace RcclUnitTesting
   int const UT_SINGLE_PROCESS = (1<<0);
   int const UT_MULTI_PROCESS  = (1<<1);
 
-  int getArchInfo(bool *isRightArch)
+  int getArchInfo(bool *isRightArch,  const char *gfx)
   {
     // Prepare parent->child pipe
     int pipefd[2];
@@ -25,7 +25,7 @@ namespace RcclUnitTesting
     }
     pid_t pid = fork();
     if (0 == pid) {
-      bool isGfx94 = false;
+      bool isGfxTest = false;
       int dev;
       hipGetDeviceCount(&dev);
       for (int deviceId = 0; deviceId < dev; deviceId++) {
@@ -34,14 +34,14 @@ namespace RcclUnitTesting
         hipGetDeviceProperties(&devProp, deviceId);
         char *gcnArchNameToken = strtok(devProp.gcnArchName, ":");
         strcpy(gcn, gcnArchNameToken);
-        if(std::strncmp("gfx94", gcn, 5) == 0) {
-          isGfx94 = true;
+        if(std::strncmp(gfx, gcn, 5) == 0) {
+          isGfxTest = true;
         } else {
-          isGfx94 = false;
+          isGfxTest = false;
           break;
         }
       }
-      if (write(pipefd[1], &isGfx94, sizeof(isGfx94)) != sizeof(isGfx94)) return TEST_FAIL;
+      if (write(pipefd[1], &isGfxTest, sizeof(isGfxTest)) != sizeof(isGfxTest)) return TEST_FAIL;
       close(pipefd[0]);
       close(pipefd[1]);
       exit(EXIT_SUCCESS);
@@ -95,7 +95,9 @@ namespace RcclUnitTesting
     numDetectedGpus = 0;
     getDeviceCount(&numDetectedGpus);
     isGfx94 = false;
-    getArchInfo(&isGfx94);
+    getArchInfo(&isGfx94, "gfx94");
+    isGfx12 = false;
+    getArchInfo(&isGfx12, "gfx12");
 
     showNames      = GetEnvVar("UT_SHOW_NAMES"  , 1);
     minGpus        = GetEnvVar("UT_MIN_GPUS"    , 2);

test/common/EnvVars.hpp

@@ -31,6 +31,7 @@ namespace RcclUnitTesting
     int  timeoutUs;      // Set timeout for child in microseconds  [UT_TIMEOUT_US]
     bool useMultithreading; // Multi-thread single-process ranks   [UT_MULTITHREAD]
     bool isGfx94;        // Detects if architecture is gfx94
+    bool isGfx12;        // Detects if architecture is gfx12
 
     // Constructor that parses and collects environment variables
     EnvVars();