Merge branch 'master' of github.com:IntelPython/pydppl

Author: diptorupd

conda-recipe/build.sh

@@ -41,6 +41,7 @@ export OpenCL_LIBDIR=${DPCPP_ROOT}/lib
 export DPPL_ONEAPI_INTERFACE_LIBDIR=${PREFIX}/lib
 export DPPL_ONEAPI_INTERFACE_INCLDIR=${PREFIX}/include
 
+
 # FIXME: How to pass this using setup.py? This flags is needed when
 # dpcpp compiles the generated cpp file.
 export CFLAGS="-fPIC -O3 ${CFLAGS}"

dppl/ocldrv.py

@@ -429,6 +429,18 @@ def create_device_array(self, array):
 
         return DeviceArray(self._env_ptr, array)
 
+    def device_support_int64_atomics(self):
+        ''' Returns True current device supports 64-bit int atomic operations
+        '''
+
+        return self._env_ptr.support_int64_atomics
+
+    def device_support_float64_atomics(self):
+        ''' Returns True if current device supports 64-bit float atomic operations
+        '''
+
+        return self._env_ptr.support_float64_atomics
+
     def get_context_ptr(self):
         ''' Returns a cdata wrapper for the OpenCL cl_context object.
         '''
@@ -541,6 +553,7 @@ def has_gpu_device(self):
 
         return Runtime._gpu_device is not None
 
+
     def get_cpu_device(self):
         ''' Returns a cdata wrapper for the first available OpenCL
         CPU context.

oneapi_wrapper/include/dppl_opencl_interface.h

@@ -48,6 +48,8 @@ struct dp_env
     void *queue;
     unsigned int max_work_item_dims;
     size_t max_work_group_size;
+    int support_int64_atomics;
+    int support_float64_atomics;
     int (*dump_fn) (void *);
 };
 

oneapi_wrapper/source/dppl_opencl_interface.c

@@ -25,6 +25,7 @@
 //===----------------------------------------------------------------------===//
 #include "dppl_opencl_interface.h"
 #include "error_check_macros.h"
+#include <string.h>
 #include <assert.h>
 #include <CL/cl.h>  /* OpenCL headers */
 
@@ -277,6 +278,49 @@ static int get_first_device (cl_platform_id* platforms,
     return DP_GLUE_FAILURE;
 }
 
+static int support_int64_atomics(cl_device_id *device)
+{
+
+    cl_int err;
+    size_t size;
+    char *value;
+
+    clGetDeviceInfo(*device, CL_DEVICE_EXTENSIONS, 0, NULL, &size);
+    if (err != CL_SUCCESS ) {
+        printf("Unable to obtain device info for param\n");
+        return DP_GLUE_FAILURE;
+    }
+    value = (char*) malloc(size);
+    clGetDeviceInfo(*device, CL_DEVICE_EXTENSIONS, size, value, NULL);
+
+    if(strstr(value, "cl_khr_int64_base_atomics") != NULL) {
+        return DP_GLUE_SUCCESS;
+    } else {
+        return DP_GLUE_FAILURE;
+    }
+}
+
+static int support_float64_atomics(cl_device_id *device)
+{
+
+    cl_int err;
+    size_t size;
+    char *value;
+
+    clGetDeviceInfo(*device, CL_DEVICE_EXTENSIONS, 0, NULL, &size);
+    if (err != CL_SUCCESS ) {
+        printf("Unable to obtain device info for param\n");
+        return DP_GLUE_FAILURE;
+    }
+    value = (char*) malloc(size);
+    clGetDeviceInfo(*device, CL_DEVICE_EXTENSIONS, size, value, NULL);
+
+    if(strstr(value, "cl_khr_fp64") != NULL) {
+        return DP_GLUE_SUCCESS;
+    } else {
+        return DP_GLUE_FAILURE;
+    }
+}
 
 /*!
  *
@@ -332,6 +376,19 @@ static int create_dp_env_t (cl_platform_id* platforms,
 
     env->device = *device;
     env ->dump_fn = dump_device_info;
+
+    if (DP_GLUE_SUCCESS == support_int64_atomics(device)) {
+        env->support_int64_atomics = 1;
+    } else {
+        env->support_int64_atomics = 0;
+    }
+
+    if (DP_GLUE_SUCCESS == support_float64_atomics(device)) {
+        env->support_float64_atomics = 1;
+    } else {
+        env->support_float64_atomics = 0;
+    }
+
     free(device);
     *env_t_ptr = env;