Pass buffer location property through OpenCL runtime for USM shared and host allocation calls.

Author: bsyrowik

include/acl_hal.h

@@ -230,7 +230,8 @@ typedef struct {
 
   /// Allocates USM host memory
   void *(*host_alloc)(const std::vector<cl_device_id> devices, size_t size,
-                      size_t alignment, int *error);
+                      size_t alignment, mem_properties_t *properties,
+                      int *error);
 
   /// Frees allocated memory by the MMD
   int (*free)(cl_context context, void *ptr);

src/acl_hal_mmd.cpp

@@ -65,7 +65,8 @@ void *acl_hal_mmd_shared_alloc(cl_device_id device, size_t size,
                                size_t alignment, mem_properties_t *properties,
                                int *error);
 void *acl_hal_mmd_host_alloc(const std::vector<cl_device_id> devices,
-                             size_t size, size_t alignment, int *error);
+                             size_t size, size_t alignment,
+                             mem_properties_t *properties, int *error);
 int acl_hal_mmd_free(cl_context context, void *mem);
 int acl_hal_mmd_try_devices(cl_uint num_devices, const cl_device_id *devices,
                             cl_platform_id platform);
@@ -2519,7 +2520,7 @@ void *acl_hal_mmd_legacy_shared_alloc(cl_context context, size_t size,
     int error = 0;
     std::vector<cl_device_id> devices = std::vector<cl_device_id>(
         context->device, context->device + context->num_devices);
-    void *mem = acl_hal_mmd_host_alloc(devices, size, 0, &error);
+    void *mem = acl_hal_mmd_host_alloc(devices, size, 0, nullptr, &error);
     device_ptr_out = static_cast<unsigned long long *>(mem);
     if (error) {
       switch (error) {
@@ -2728,7 +2729,8 @@ void *acl_hal_mmd_shared_alloc(cl_device_id device, size_t size,
 }
 
 void *acl_hal_mmd_host_alloc(const std::vector<cl_device_id> devices,
-                             size_t size, size_t alignment, int *error) {
+                             size_t size, size_t alignment,
+                             mem_properties_t *properties, int *error) {
   // Note we do not support devices in the same context with different MMDs
   // Safe to get the mmd handle from first device
   void *result = NULL;
@@ -2749,8 +2751,9 @@ void *acl_hal_mmd_host_alloc(const std::vector<cl_device_id> devices,
     handles[i] = device_info[physical_device_id].handle;
   }
 
-  result = dispatch->aocl_mmd_host_alloc(handles, devices.size(), size,
-                                         alignment, nullptr, error);
+  result = dispatch->aocl_mmd_host_alloc(
+      handles, devices.size(), size, alignment,
+      (aocl_mmd_mem_properties_t *)properties, error);
   acl_delete_arr(handles);
 
   if (error) {

src/acl_usm.cpp

@@ -3,9 +3,11 @@
 
 // System headers.
 #include <algorithm>
+#include <array>
 #include <cassert>
 #include <cstdio>
 #include <iostream>
+#include <optional>
 #include <sstream>
 #include <string>
 #include <unordered_set>
@@ -28,6 +30,8 @@
 #include <acl_usm.h>
 #include <acl_util.h>
 
+#include <MMD/aocl_mmd.h>
+
 #ifdef __GNUC__
 #pragma GCC visibility push(protected)
 #endif
@@ -99,11 +103,15 @@ CL_API_ENTRY void *CL_API_CALL clHostMemAllocINTEL(
   // Iterate over properties.
   // The end of the properties list is specified with a zero.
   cl_mem_alloc_flags_intel alloc_flags = 0;
+  std::optional<cl_uint> mem_id;
   while (properties != NULL && *properties != 0) {
     switch (*properties) {
     case CL_MEM_ALLOC_FLAGS_INTEL: {
       alloc_flags = *(properties + 1);
     } break;
+    case CL_MEM_ALLOC_BUFFER_LOCATION_INTEL: {
+      mem_id = (cl_uint) * (properties + 1);
+    } break;
     default: {
       UNLOCK_BAIL_INFO(CL_INVALID_PROPERTY, context, "Invalid properties");
     }
@@ -138,6 +146,16 @@ CL_API_ENTRY void *CL_API_CALL clHostMemAllocINTEL(
   }
 
   if (acl_get_hal()->host_alloc) {
+    std::array<mem_properties_t, 3> mmd_properties;
+    {
+      auto mmd_properties_it = mmd_properties.begin();
+      if (mem_id) {
+        *mmd_properties_it++ = AOCL_MMD_MEM_PROPERTIES_BUFFER_LOCATION;
+        *mmd_properties_it++ = *mem_id;
+      }
+      *mmd_properties_it++ = 0;
+    }
+
     acl_usm_allocation_t *usm_alloc =
         (acl_usm_allocation_t *)acl_malloc(sizeof(acl_usm_allocation_t));
 
@@ -146,7 +164,8 @@ CL_API_ENTRY void *CL_API_CALL clHostMemAllocINTEL(
     }
 
     int error = 0;
-    void *mem = acl_get_hal()->host_alloc(devices, size, alignment, &error);
+    void *mem = acl_get_hal()->host_alloc(devices, size, alignment,
+                                          mmd_properties.data(), &error);
     if (error) {
       acl_free(usm_alloc);
       switch (error) {
@@ -380,6 +399,7 @@ clSharedMemAllocINTEL(cl_context context, cl_device_id device,
   // The end of the properties list is specified with a zero.
   cl_mem_alloc_flags_intel alloc_flags = 0;
   std::unordered_set<unsigned long long> seen_flags;
+  std::optional<cl_uint> mem_id;
   while (properties != NULL && *properties != 0) {
     switch (*properties) {
     case CL_MEM_ALLOC_FLAGS_INTEL: {
@@ -396,6 +416,14 @@ clSharedMemAllocINTEL(cl_context context, cl_device_id device,
       }
       alloc_flags = *(properties + 1);
     } break;
+    case CL_MEM_ALLOC_BUFFER_LOCATION_INTEL: {
+      if (seen_flags.insert(CL_MEM_ALLOC_BUFFER_LOCATION_INTEL).second ==
+          false) {
+        UNLOCK_BAIL_INFO(CL_INVALID_PROPERTY, context,
+                         "Property specified multiple times");
+      }
+      mem_id = (cl_uint) * (properties + 1);
+    } break;
     default: {
       UNLOCK_BAIL_INFO(CL_INVALID_PROPERTY, context, "Invalid properties");
     }
@@ -404,6 +432,16 @@ clSharedMemAllocINTEL(cl_context context, cl_device_id device,
   }
 
   if (acl_get_hal()->shared_alloc) {
+    std::array<mem_properties_t, 3> mmd_properties;
+    {
+      auto mmd_properties_it = mmd_properties.begin();
+      if (mem_id) {
+        *mmd_properties_it++ = AOCL_MMD_MEM_PROPERTIES_BUFFER_LOCATION;
+        *mmd_properties_it++ = *mem_id;
+      }
+      *mmd_properties_it++ = 0;
+    }
+
     acl_usm_allocation_t *usm_alloc =
         (acl_usm_allocation_t *)acl_malloc(sizeof(acl_usm_allocation_t));
 
@@ -412,8 +450,8 @@ clSharedMemAllocINTEL(cl_context context, cl_device_id device,
     }
 
     int error;
-    void *mem =
-        acl_get_hal()->shared_alloc(device, size, alignment, nullptr, &error);
+    void *mem = acl_get_hal()->shared_alloc(device, size, alignment,
+                                            mmd_properties.data(), &error);
     if (mem == NULL) {
       acl_free(usm_alloc);
       switch (error) {

test/acl_hal_test.cpp

@@ -91,7 +91,8 @@ void *acl_test_hal_shared_alloc(cl_device_id device, size_t size,
                                 size_t alignment, mem_properties_t *properties,
                                 int *error);
 void *acl_test_hal_host_alloc(const std::vector<cl_device_id> devices,
-                              size_t size, size_t alignment, int *error);
+                              size_t size, size_t alignment,
+                              mem_properties_t *properties, int *error);
 int acl_test_hal_free(cl_context context, void *ptr);
 
 static acl_event_update_callback acltest_hal_event_callback = NULL;
@@ -662,7 +663,8 @@ void *acl_test_hal_shared_alloc(cl_device_id device, size_t size,
 }
 
 void *acl_test_hal_host_alloc(const std::vector<cl_device_id> devices,
-                              size_t size, size_t alignment, int *error) {
+                              size_t size, size_t alignment,
+                              mem_properties_t *properties, int *error) {
   size = size;
   alignment = alignment;
   error = error;