Make runtime return true device capabilities

Author: sophimao

src/acl_device.cpp

@@ -2,7 +2,6 @@
 // SPDX-License-Identifier: BSD-3-Clause
 
 // System headers.
-#include <algorithm>
 #include <assert.h>
 #include <string.h>
 
@@ -279,23 +278,37 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
     // Cut by 2 again, see comment for CL_DEVICE_GLOBAL_MEM_SIZE
     size /= 2;
 #else
-    // Constant memory is global memory.
+    // Return the maximum size of a single allocation to the constant memory
+    // (i.e., global memory)
     cl_ulong size = 0;
     for (unsigned gmem_idx = 0;
          gmem_idx < device->def.autodiscovery_def.num_global_mem_systems;
          gmem_idx++) {
       if (device->def.autodiscovery_def.global_mem_defs[gmem_idx].type ==
           ACL_GLOBAL_MEM_DEVICE_PRIVATE) {
-        size += ACL_RANGE_SIZE(
-            device->def.autodiscovery_def.global_mem_defs[gmem_idx].range);
+        cl_ulong curr_size = 0;
+        // TODO: investigate if ACL_MEM_ALIGN of 0x400 is still required to
+        //       perform device allocations to memory with 0 starting address
+        if (device->def.autodiscovery_def.global_mem_defs[gmem_idx]
+                .allocation_type &
+            ACL_GLOBAL_MEM_DEVICE_ALLOCATION) {
+          curr_size = ACL_RANGE_SIZE(
+              device->def.autodiscovery_def.global_mem_defs[gmem_idx]
+                  .get_usable_range());
+        } else {
+          curr_size = ACL_RANGE_SIZE(
+              device->def.autodiscovery_def.global_mem_defs[gmem_idx].range);
+        }
+        if (curr_size > size) {
+          size = curr_size;
+        }
       }
     }
-    // Note that OpenCL 1.2 specifies the minimum value for devices not of
-    // type CL_DEVICE_TYPE_CUSTOM to be 64KB, however, the OpenCL conformance
-    // test api:test_min_max_constant_buffer_size expects to allocate two
-    // buffers of the size returned here, so following the spec may result
-    // in conformance test failures.
-    size = std::max(size, (cl_ulong)64 * 1024);
+    // Note: devices not of type CL_DEVICE_TYPE_CUSTOM and conformant
+    // to OpenCL 1.2 spec will return size at least of 64KB here
+    // TODO: the OpenCL conformance test api:test_min_max_constant_buffer_size
+    //       expects to allocate two buffers of the size returned, need to
+    //       confirm conformance here
 #endif
     RESULT_ULONG(size);
   } break;
@@ -322,15 +335,12 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
       size = size / 8;
     }
 #else
-    cl_ulong global_mem_size = 0;
     cl_ulong size = 0;
     for (unsigned gmem_idx = 0;
          gmem_idx < device->def.autodiscovery_def.num_global_mem_systems;
          gmem_idx++) {
       if (device->def.autodiscovery_def.global_mem_defs[gmem_idx].type ==
           ACL_GLOBAL_MEM_DEVICE_PRIVATE) {
-        global_mem_size += ACL_RANGE_SIZE(
-            device->def.autodiscovery_def.global_mem_defs[gmem_idx].range);
         cl_ulong curr_size = 0;
         // TODO: investigate if ACL_MEM_ALIGN of 0x400 is still required to
         //       perform device allocations to memory with 0 starting address
@@ -349,10 +359,9 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
         }
       }
     }
-    // OpenCL 1.2: min value = max(CL_DEVICE_GLOBAL_MEM_SIZE/4, 1*1024*1024)
-    // for devices that are not of type CL_DEVICE_TYPE_CUSTOM
-    size = std::max(size,
-                    std::max(global_mem_size / 4, (cl_ulong)1 * 1024 * 1024));
+    // Note: devices not of type CL_DEVICE_TYPE_CUSTOM and
+    // conformant to OpenCL 1.2 spec will return size at least of
+    // max(CL_DEVICE_GLOBAL_MEM_SIZE/4, 1*1024*1024) here
 #endif
     RESULT_ULONG(size);
   } break;