add new backend type: c-ocl_nvidia (#194)

backends/c-ocl_amdgpu/include/backend.hpp

@@ -2,6 +2,7 @@
 // Licensed under the MIT License.
 
 //; eval_flags(c-ocl_amdgpu): -I/opt/rocm/opencl/include -L/opt/rocm/opencl/lib -lOpenCL -DCL_TARGET_OPENCL_VERSION=120
+//; eval_flags(c-ocl_nvidia): -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -lOpenCL
 
 #include <CL/cl.h>
 
@@ -10,6 +11,7 @@ namespace ab {
   static cl_command_queue cmdqueue;
   static cl_device_id device_id;
   static cl_int stat;
+  static size_t max_work_group_size;
 
   void init(int dev) {
     cl_uint num_dev;
@@ -28,7 +30,6 @@ namespace ab {
     cmdqueue = clCreateCommandQueue(context, device_id, 0, &stat), CHECK_OK(stat == 0);
 
     std::vector<char> dev_name(1024);
-    size_t max_work_group_size = 0;
     CHECK_OK(0 == clGetDeviceInfo(device_id, CL_DEVICE_NAME, dev_name.size(), dev_name.data(), NULL));
     CHECK_OK(0 == clGetDeviceInfo(device_id, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(max_work_group_size), &max_work_group_size, NULL));
     fprintf(stderr, "    (OCL_INFO: OCL Device Name = %s [max_work_groups: %zd])\n", dev_name.data(), max_work_group_size);
@@ -74,6 +75,8 @@ namespace ab {
     };
     size_t lx = query("get_local_id(0)"), ly = query("get_local_id(1)"), lz = query("get_local_id(2)");
     size_t gx = query("get_group_id(0)"), gy = query("get_group_id(1)"), gz = query("get_group_id(2)");
+
+    CHECK_OK(lx * ly * lz <= max_work_group_size);
     return { kernel, (void*)(gx * lx), (void*)(gy * ly), (void*)(gz * lz), (void*)lx, (void*)ly, (void*)lz };
   }
 

backends/c-ocl_nvidia

@@ -0,0 +1 @@
+c-ocl_amdgpu

graph_evaluator/run_graph.cpp

@@ -74,7 +74,7 @@ int main(int argc, char** argv)
       printf("\n- K/%d: %.10e\n", i, digest);
     }
 
-    {
+    do {
       auto x = ab::recordTime();
       gm.compute(global_args.data());
       auto y = ab::recordTime();
@@ -82,8 +82,10 @@ int main(int argc, char** argv)
 
       double tpr = ab::convertToElapsedTime(x, y);
       const char *expected_timeout = getenv("EXPECTED_TIMEOUT");
-      if (expected_timeout && *expected_timeout && tpr > std::atof(expected_timeout))
-        throw std::runtime_error(("Time limit exceeded: " + std::to_string(tpr) + " v.s. (expected) " + expected_timeout).c_str());
+      if (expected_timeout && *expected_timeout && tpr > std::atof(expected_timeout)) {
+        printf("\n- TPR: %g\n", tpr);
+        break;
+      }
 
       int num_runs = (int)std::max(1LU, std::min(10000LU, (unsigned long)(1.0 / tpr)));
       tpr = 0.0f;
@@ -93,7 +95,7 @@ int main(int argc, char** argv)
       y = ab::recordTime();
       tpr = ab::convertToElapsedTime(x, y) / num_runs;
       printf("\n- TPR: %g\n", tpr);
-    }
+    } while (0);
 
     ab::finalize();
     return 0;

hardware/OCL_DEFAULT.cfg

@@ -1,7 +1,7 @@
-MaxThreadsPerBlock: 256
-MaxBlockDimX: 256
-MaxBlockDimY: 256
-MaxBlockDimZ: 256
+MaxThreadsPerBlock: 1024
+MaxBlockDimX: 1024
+MaxBlockDimY: 1024
+MaxBlockDimZ: 1024
 MaxSharedMemoryPerBlock: 65536
 WarpSize: 32
 ClockRate: 1000000