adding multi-GPU test

test/MatrixTranspose_test/Makefile

@@ -11,24 +11,24 @@ ifeq (,$(HIP_PATH))
 endif
 
 HIPCC=$(HIP_PATH)/bin/hipcc
-
 TARGET=hcc
 
-
-
+EXECUTABLE=./MatrixTranspose
+OBJECTS = MatrixTranspose.o
 FLAGS =-g -I$(ROOT_PATH) -I$(ROOT_PATH)/inc -I${HSA_KMT_INC_PATH} -DLOCAL_BUILD=1 -DHIP_VDI=${HIP_VDI} -DITERATIONS=$(ITERATIONS)
+
 ifeq ($(C_TEST), 1)
 	COMP=gcc
 	SOURCES = MatrixTranspose.c
 	FLAGS += -DHIP_TEST=0 -D__HIP_PLATFORM_HCC__=1 -I/opt/rocm/hcc/include
-	EXECUTABLE=./MatrixTranspose_ctest
 else
 	COMP=$(HIPCC)
-	FLAGS += -DHIP_TEST=1
 	SOURCES = MatrixTranspose.cpp
-	EXECUTABLE=./MatrixTranspose
+	FLAGS += -DHIP_TEST=1
+endif
+ifeq ($(MGPU_TEST), 1)
+	FLAGS += -DMGPU_TEST=1
 endif
-OBJECTS = MatrixTranspose.o
 
 .PHONY: test
 

test/MatrixTranspose_test/MatrixTranspose.cpp

@@ -84,21 +84,33 @@ int main() {
     float* gpuMatrix;
     float* gpuTransposeMatrix;
 
-#if HIP_TEST
-    hipDeviceProp_t devProp;
-    HIP_CALL(hipGetDeviceProperties(&devProp, 0));
-
-    printf("Device name %s\n", devProp.name);
-#endif
-
     int i;
     int errors;
 
     init_tracing();
 
+#if HIP_TEST
+    int gpuCount = 1;
+#if MGPU_TEST
+    hipGetDeviceCount(&gpuCount);
+    printf("Number of GPUs: %d\n", gpuCount);
+#endif
+    iterations *= gpuCount;
+#endif
+
     while (iterations-- > 0) {
     start_tracing();
 
+#if HIP_TEST
+    // set GPU
+    const int devIndex = iterations % gpuCount;
+    hipSetDevice(devIndex);
+
+    hipDeviceProp_t devProp;
+    HIP_CALL(hipGetDeviceProperties(&devProp, 0));
+    printf("Device %d name: %s\n", devIndex, devProp.name);
+#endif
+
     Matrix = (float*)malloc(NUM * sizeof(float));
     TransposeMatrix = (float*)malloc(NUM * sizeof(float));
     cpuTransposeMatrix = (float*)malloc(NUM * sizeof(float));