Automatically set minimum tensor core compute capability 8.0 and adde…

…d error message if device is below.

Makefile

@@ -27,7 +27,11 @@
 OVERLAP = ON
 
 ifeq ($(DEVICE), $(filter $(DEVICE),GPU CUDA))
-TARGETS_SUPPORTED := $(shell ./test_cuda.sh nvcc "$(GPU_INCLUDE_PATH)" "$(GPU_LIBRARY_PATH)" "$(TARGETS)" "$(DEVICE)")
+MIN_COMPUTE:=50
+ifeq ($(TENSOR), ON)
+MIN_COMPUTE:=80
+endif
+TARGETS_SUPPORTED := $(shell ./test_cuda.sh nvcc "$(GPU_INCLUDE_PATH)" "$(GPU_LIBRARY_PATH)" "$(TARGETS)" "$(MIN_COMPUTE)")
 # if user specifies DEVICE=GPU the test result determines wether CUDA will be used or not
 ifeq ($(TARGETS_SUPPORTED),)
 ifeq ($(DEVICE),CUDA)

host/src/performdocking.cpp

@@ -450,6 +450,13 @@ void setup_gpu_for_docking(
 	cudaDeviceProp props;
 	RTERROR(cudaGetDevice(&(cData.devnum)),"ERROR in cudaGetDevice:");
 	RTERROR(cudaGetDeviceProperties(&props,cData.devnum),"ERROR in cudaGetDeviceProperties:");
+#ifdef USE_NVTENSOR
+	if(props.major < 8){
+		printf("Error: Compute capability 8.0 or higher is needed for tensor core sum reductions.\n");
+		printf("       Available device %s has compute capability %d.%d.\n", props.name, props.major, props.minor);
+		exit(-1);
+	}
+#endif
 	tData.device_name = (char*) malloc(strlen(props.name)+32); // make sure array is large enough to hold device number text too
 	strcpy(tData.device_name, props.name);
 	if(gpuCount>1) snprintf(&tData.device_name[strlen(props.name)], strlen(props.name)+32, " (#%d / %d)",cData.devnum+1,gpuCount);

test_cuda.sh

@@ -26,7 +26,7 @@ if [[ "$4" != "" ]]; then
 	done
 	TARGETS="$4"
 else
-	TARGETS=`awk -F'_' '{ if(\$2>50) print \$2 }' <<< "$TARGETS_SUPPORTED" | tr "\n" " "`
+	TARGETS=`awk -F'_' "{ if(\\$2>=$5) print \\$2 }" <<< "$TARGETS_SUPPORTED" | tr "\n" " "`
 fi
 printf "Compiling for targets: %s\n" "$TARGETS" >&2
 cd "$script_dir"