Add buffer location property

Sycl runtime calls clEnqueueWriteBuffer before clSetKernelArgs, this cause the buffer to be allocated on the device before knowing the right place of allocating the memory. As a result, later when kernel get invoked, the memory has to be copied from device's default global memory to the buffer location specified in kernel. This is an additional memory copy operation.

This extension does not interfer with the other way of setting buffer location (i.e through clSetKernelArgs). This property exist for integration with sycl runtime, not for pure opencl user to use. If opencl user wish to use this property, they have to make sure the buffer location passed into clCreateBufferWithPropertyINTEL has to match the one defined in kernel function interface, otherwise the extra memory copy issue will remain.

When resizing reserved allocation, we now have the information to allocate minimum amount of space required according to the property passed in.

Author: sherry-yuan

include/CL/cl_ext_intel.h

@@ -441,10 +441,11 @@ typedef cl_bitfield cl_mem_alloc_flags_intel;
 typedef cl_uint cl_mem_info_intel;
 
 /* cl_mem_alloc_info_intel */
-#define CL_MEM_ALLOC_TYPE_INTEL         0x419A
-#define CL_MEM_ALLOC_BASE_PTR_INTEL     0x419B
-#define CL_MEM_ALLOC_SIZE_INTEL         0x419C
-#define CL_MEM_ALLOC_DEVICE_INTEL       0x419D
+#define CL_MEM_ALLOC_TYPE_INTEL               0x419A
+#define CL_MEM_ALLOC_BASE_PTR_INTEL           0x419B
+#define CL_MEM_ALLOC_SIZE_INTEL               0x419C
+#define CL_MEM_ALLOC_DEVICE_INTEL             0x419D
+#define CL_MEM_ALLOC_BUFFER_LOCATION_INTEL    0x419E
 /* Enum values 0x419E-0x419F are reserved for future queries. */
 
 typedef cl_uint cl_unified_shared_memory_type_intel;

src/acl_mem.cpp

@@ -416,6 +416,7 @@ CL_API_ENTRY cl_mem clCreateBufferWithPropertiesINTEL(
   cl_bool context_has_device_with_physical_mem;
   unsigned int idevice;
   cl_uint bank_id = 0;
+  unsigned int tmp_mem_id = 0;
   acl_lock();
 
 #ifdef MEM_DEBUG_MSG
@@ -431,6 +432,9 @@ CL_API_ENTRY cl_mem clCreateBufferWithPropertiesINTEL(
       }
       bank_id = (cl_uint) * (properties + 1);
     } break;
+    case CL_MEM_ALLOC_BUFFER_LOCATION_INTEL: {
+      tmp_mem_id = (unsigned int)*(properties + 1);
+    } break;
     default: {
       UNLOCK_BAIL_INFO(CL_INVALID_DEVICE, context, "Invalid properties");
     }
@@ -553,6 +557,7 @@ CL_API_ENTRY cl_mem clCreateBufferWithPropertiesINTEL(
     UNLOCK_BAIL_INFO(CL_OUT_OF_HOST_MEMORY, context,
                      "Could not allocate a cl_mem object");
   }
+  mem->mem_id = tmp_mem_id;
 
   mem->block_allocation = new_block;
   mem->block_allocation->mem_obj = mem;
@@ -784,7 +789,6 @@ CL_API_ENTRY cl_mem clCreateBufferWithPropertiesINTEL(
     mem->context = context;
     mem->flags = flags;
     mem->size = size;
-    mem->mem_id = 0;
 
     mem->bank_id = 0;
     if (is_SOC_device()) {
@@ -1245,7 +1249,7 @@ CL_API_ENTRY cl_mem CL_API_CALL clCreateSubBufferIntelFPGA(
 
     mem->context = context;
     mem->flags = sub_flags;
-    mem->mem_id = 0;
+    mem->mem_id = buffer->mem_id;
 
     if (is_SOC_device()) {
       // HPS DDR is system managed for SoC.
@@ -4408,6 +4412,10 @@ void acl_resize_reserved_allocations_for_device(cl_mem mem,
   unsigned int num_global_mem_systems =
       def.autodiscovery_def.num_global_mem_systems;
 
+  // When we don't know how many memory systems will exist
+  // Load as much as needed.
+  num_global_mem_systems = std::max(num_global_mem_systems, mem->mem_id + 1);
+
   // For the simulation flow we don't know how many memory systems will exist
   // until we load the .aocx, which may not happen until somewhat later.
   // Reserving space is quite cheap, so reserve space for many memory systems.

src/acl_usm.cpp

@@ -255,8 +255,8 @@ clDeviceMemAllocINTEL(cl_context context, cl_device_id device,
   cl_int status;
 
   // Use cl_mem for convenience
-  cl_mem usm_device_buffer =
-      clCreateBufferIntelFPGA(context, CL_MEM_READ_WRITE, size, NULL, &status);
+  cl_mem usm_device_buffer = clCreateBufferWithPropertiesINTEL(
+      context, NULL, CL_MEM_READ_WRITE, size, NULL, &status);
   if (status != CL_SUCCESS) {
     UNLOCK_BAIL_INFO(status, context, "Failed to allocate device memory");
   }

test/acl_mem_test.cpp

@@ -2800,6 +2800,27 @@ TEST(acl_mem, case_205751_overlapping_alloc) {
   CHECK_EQUAL(CL_SUCCESS, clReleaseMemObject(c));
 }
 
+TEST(acl_mem, buffer_location_property) {
+  ACL_LOCKED(acl_print_debug_msg("begin buffer_location_property\n"));
+  // Test assumes more than 1 global memory space
+  // Allocate a small buffer (a), then try to allocate two buffers (b, c) of
+  // size bank_size.  Expect the second allocation to fail.
+  cl_mem a;
+  cl_int status = CL_SUCCESS;
+  size_t total_size = ACL_RANGE_SIZE(
+      m_device[0]->def.autodiscovery_def.global_mem_defs[0].range);
+  size_t bank_size = total_size / 2;
+  size_t small_size = bank_size / 1024;
+
+  cl_mem_properties_intel props[] = {CL_MEM_ALLOC_BUFFER_LOCATION_INTEL, 0, 0};
+  a = clCreateBufferWithPropertiesINTEL(m_context, props, 0, bank_size, 0,
+                                        &status);
+  CHECK_EQUAL(CL_SUCCESS, status);
+  ACL_LOCKED(CHECK_EQUAL(acl_bind_buffer_to_device(m_cq->device, a), 1));
+
+  CHECK_EQUAL(CL_SUCCESS, clReleaseMemObject(a));
+}
+
 MT_TEST(acl_mem, map_buf_bad_flags) {
   ACL_LOCKED(acl_print_debug_msg("begin buf_bad_flags\n"));
   cl_int status = CL_SUCCESS;