Support arbitary BL for accessor

Author: zibaiwan

include/acl.h

@@ -381,6 +381,8 @@ typedef struct {
 
     return result;
   }
+
+  std::string id = "-";
 } acl_system_global_mem_def_t;
 
 // Our allocator is optimized for simplicity, and for the case where there

src/acl_auto_configure.cpp

@@ -278,6 +278,8 @@ read_global_mem_defs(const std::string &config_str,
 
   for (auto i = 0U; result && (i < num_global_mem_systems); i++) {
     std::string gmem_name;
+    std::string gmem_id = "-"; // This should be the default value even if the
+                               // auto-discovery string doesn't have this field.
     // read total number of fields in global_memories
     int total_fields_global_memories = 0;
     result = read_int_counters(config_str, curr_pos,
@@ -403,6 +405,12 @@ read_global_mem_defs(const std::string &config_str,
           can_access.push_back(temp);
         }
       }
+
+      // read global memory id field. If it doesn't exist, then the value is "-"
+      // It's a new field introduced from 2024.2
+      if (result && counters.back() > 0) {
+        result = read_string_counters(config_str, curr_pos, gmem_id, counters);
+      }
     }
 
     if (result) {
@@ -428,6 +436,7 @@ read_global_mem_defs(const std::string &config_str,
       global_mem_defs[i].allocation_type = allocation_type;
       global_mem_defs[i].primary_interface = primary_interface;
       global_mem_defs[i].can_access_list = can_access;
+      global_mem_defs[i].id = gmem_id;
     }
 
     // forward compatibility: bypassing remaining fields at the end of global

src/acl_mem.cpp

@@ -437,6 +437,64 @@ CL_API_ENTRY cl_mem clCreateBufferWithPropertiesINTEL(
     } break;
     case CL_MEM_ALLOC_BUFFER_LOCATION_INTEL: {
       tmp_mem_id = (cl_uint) * (properties + 1);
+
+      // In FullSystem flow, buffer location is always the index of the global
+      // memories. Therefore, there is no additional handling needed for FS.
+
+      // However, in SYCL_HLS(IPA) flow, the user passed buffer_location<id>
+      // maps to the global memory with the same id. Runtime needs to find the
+      // correct index of the global memory with that id. The id filed is
+      // introduced in 2024.2 in auto-discovery string. This change is for
+      // accessor only and the USM buffer location change is done in the
+      // simulator.
+
+      // Runtime needs to determine whether it's FS or SYCL_HLS first by
+      // checking if the global memory id exist or not. If exists, then it's
+      // SYCL_HLS flow.
+      bool is_SYCL_HLS = false;
+
+      // We document the limitation here:
+      // https://www.intel.com/content/www/us/en/docs/oneapi-fpga-add-on/
+      // developer-guide/2024-0/targeting-multiple-homogeneous-fpga-devices.html
+      // All FPGA devices used must be of the same FPGA card (same -Xstarget
+      // target). There might be some workaround supporting multi-device with
+      // different boards but they are for FS, not for SYCL_HLS.
+
+      // Therefore, we can safely assume all devices have the same
+      // global_mem_defs in SYCL_HLS flow as of 2024.2. So we can just check
+      // acl_platform.device[0].
+
+      auto global_mem_defs =
+          acl_platform.device[0].def.autodiscovery_def.global_mem_defs;
+
+      for (const auto &global_mem_def : global_mem_defs) {
+        if (global_mem_def.id != "-") {
+          is_SYCL_HLS = true;
+          break;
+        }
+      }
+
+      if (is_SYCL_HLS) {
+        // find the correct index in the global memory
+        long index = -1;
+        for (auto it = begin(global_mem_defs); it != end(global_mem_defs);
+             ++it) {
+          if (stoul(it->id) == tmp_mem_id) {
+            index = it - global_mem_defs.begin();
+            break;
+          }
+        }
+
+        if (index == -1) {
+          BAIL_INFO(CL_INVALID_VALUE, context,
+                    "Invalid Buffer Location id provided");
+        }
+
+        // Update the tmp_mem_id to the corect index in the global memories
+        // vector.
+        tmp_mem_id = static_cast<cl_uint>(index);
+      }
+
     } break;
     default: {
       BAIL_INFO(CL_INVALID_DEVICE, context, "Invalid properties");

src/acl_program.cpp

@@ -2125,6 +2125,7 @@ l_device_memory_definition_copy(acl_device_def_autodiscovery_t *dest_dev,
         src_dev->global_mem_defs[idef].primary_interface;
     dest_dev->global_mem_defs[idef].can_access_list =
         src_dev->global_mem_defs[idef].can_access_list;
+    dest_dev->global_mem_defs[idef].id = src_dev->global_mem_defs[idef].id;
   }
 }
 

test/acl_auto_configure_test.cpp

@@ -1672,3 +1672,39 @@ TEST(auto_configure, sideband_mappings) {
   CHECK(devdef.sideband_signal_mappings[7].port_offset == 352);
   CHECK(devdef.sideband_signal_mappings[7].sideband_size == 32);
 }
+
+TEST(auto_configure, global_mem_id) {
+  const std::string config_str{
+      "23 46 " RANDOM_HASH " custom_ipa 0 3 "
+      " 10 1 2 1 2 2199023255552 2233382993920 4 - 0 1"   // Global memory 1
+      " 10 3 2 1 2 6597069766656 6631429505024 4 - 0 3"   // Global memory 2
+      " 10 5 2 1 2 10995116277760 11029476016128 4 - 0 5" // Global memory 3
+      " 0 0 0 0 0 1 0 0"
+      " 0 2 133 _ZTS10SimpleVAddIiE 0 256 1 0 0 1 0 1 0 13 8 2 1 8"
+      " 1 1 5 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 8"
+      " 2 1 8 1 1 3 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0"
+      " 0 0 8 2 1 8 1 1 1 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 7 0 0"
+      " 8 1 0 0 0 7 0 0 4 1 0 0 0 0 0 0 0 1 1 1 3 1 1 1 0 1 1 0 133 "
+      " _ZTS10SimpleVAddIfE 256 256 1 0 0 1 0 1 0 13 8 2 1 8 1 1 5 0"
+      " 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 8 2 1 8 1 1"
+      " 3 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 8 2 1 8"
+      " 1 1 1 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 7 0 0 8 1 0 0 0 7 0"
+      " 0 4 1 0 0 0 0 0 0 0 1 1 1 3 1 1 1 0 1 1 0"};
+
+  acl_device_def_autodiscovery_t devdef;
+  {
+    bool result;
+    std::string err_str;
+    ACL_LOCKED(result =
+                   acl_load_device_def_from_str(config_str, devdef, err_str));
+    std::cerr << err_str;
+    CHECK(result);
+  }
+
+  CHECK("1" == devdef.global_mem_defs[0].name);
+  CHECK("1" == devdef.global_mem_defs[0].id);
+  CHECK("3" == devdef.global_mem_defs[1].name);
+  CHECK("3" == devdef.global_mem_defs[1].id);
+  CHECK("5" == devdef.global_mem_defs[2].name);
+  CHECK("5" == devdef.global_mem_defs[2].id);
+}