Adding support to hsaKmtMapMemoryToGPUNodes

Change-Id: Iab6222402a43c3cd31b0efc5a316a6482986258e
Signed-off-by: Ben Goz <ben.goz@amd.com>

include/hsakmt.h

@@ -416,6 +416,21 @@ hsaKmtMapMemoryToGPU(
     HSAuint64*      AlternateVAGPU     //OUT (page-aligned)     
     );
 
+/**
+  Ensures that the memory is resident and can be accessed by GPUs
+*/
+
+HSAKMT_STATUS
+HSAKMTAPI
+hsaKmtMapMemoryToGPUNodes(
+    void*           MemoryAddress,         //IN (page-aligned)
+    HSAuint64       MemorySizeInBytes,     //IN (page-aligned)
+    HSAuint64*      AlternateVAGPU,        //OUT (page-aligned)
+     HsaMemMapFlags    MemMapFlags,               //IN
+    HSAuint64       NumberOfNodes,         //IN
+    HSAuint32*      NodeArray              //IN
+    );
+
 /**
   Releases the residency of the memory
 */

include/hsakmttypes.h

@@ -476,7 +476,7 @@ typedef struct _HsaMemMapFlags
         struct
         {
             unsigned int Reserved1      :  1; //
-            unsigned int CachePolicy    :  4; // see HSA_CACHING_TYPE
+            unsigned int CachePolicy    :  2; // see HSA_CACHING_TYPE
             unsigned int ReadOnly       :  1; // memory is not modified while mapped
             	    	    	    	      // allows migration scale-out
 	    unsigned int PageSize	    :  2; // see HSA_PAGE_SIZE, hint to use
@@ -492,7 +492,7 @@ typedef struct _HsaMemMapFlags
 						  // may trigger eviction of nonessential
 						  // data from the memory, reduces latency
 						  // “cleanup hint” only, may be ignored
-            unsigned int Reserved       : 21;
+            unsigned int Reserved       : 23;
         };
         HSAuint32 Value;
     };

src/libhsakmt.h

@@ -63,6 +63,8 @@ HSAKMT_STATUS validate_nodeid(uint32_t nodeid, uint32_t *gpu_id);
 HSAKMT_STATUS gpuid_to_nodeid(uint32_t gpu_id, uint32_t* node_id);
 uint16_t get_device_id_by_node(HSAuint32 node_id);
 uint16_t get_device_id_by_gpu_id(HSAuint32 gpu_id);
+HSAKMT_STATUS validate_nodeid_array(uint32_t **gpu_id_array,
+		uint32_t NumberOfNodes, uint32_t *NodeArray);
 
 HSAKMT_STATUS topology_sysfs_get_gpu_id(uint32_t node_id, uint32_t *gpu_id);
 HSAKMT_STATUS topology_sysfs_get_node_props(uint32_t node_id, HsaNodeProperties *props, uint32_t *gpu_id);

src/libhsakmt.ver

@@ -28,6 +28,7 @@ hsaKmtRegisterMemory;
 hsaKmtRegisterMemoryToNodes;
 hsaKmtDeregisterMemory;
 hsaKmtMapMemoryToGPU;
+hsaKmtMapMemoryToGPUNodes;
 hsaKmtUnmapMemoryToGPU;
 hsaKmtDbgRegister;
 hsaKmtDbgUnregister;

src/memory.c

@@ -212,21 +212,10 @@ hsaKmtRegisterMemoryToNodes(
 {
 	CHECK_KFD_OPEN();
 	uint32_t *gpu_id_array;
-	unsigned i;
 	HSAKMT_STATUS ret = HSAKMT_STATUS_SUCCESS;
 
-	if (NumberOfNodes == 0 || NodeArray == NULL)
-		return HSAKMT_STATUS_INVALID_PARAMETER;
-
-	/* Translate Node IDs to gpu_ids */
-	gpu_id_array = malloc(NumberOfNodes * sizeof(uint32_t));
-	if (gpu_id_array == NULL)
-		return HSAKMT_STATUS_NO_MEMORY;
-        for (i = 0; i < NumberOfNodes; i++) {
-		ret = validate_nodeid(NodeArray[i], &gpu_id_array[i]);
-		if (ret != HSAKMT_STATUS_SUCCESS)
-			break;
-	}
+	ret = validate_nodeid_array(&gpu_id_array,
+			NumberOfNodes, NodeArray);
 
 	if (ret == HSAKMT_STATUS_SUCCESS)
 		ret = fmm_register_memory(MemoryAddress, MemorySizeInBytes,
@@ -269,6 +258,36 @@ hsaKmtMapMemoryToGPU(
 		return HSAKMT_STATUS_ERROR;
 }
 
+HSAKMT_STATUS
+HSAKMTAPI
+hsaKmtMapMemoryToGPUNodes(
+    void*           MemoryAddress,         //IN (page-aligned)
+    HSAuint64       MemorySizeInBytes,     //IN (page-aligned)
+    HSAuint64*      AlternateVAGPU,        //OUT (page-aligned)
+    HsaMemMapFlags  MemMapFlags,           //IN
+    HSAuint64       NumberOfNodes,         //IN
+    HSAuint32*      NodeArray              //IN
+)
+{
+	uint32_t *gpu_id_array;
+	HSAKMT_STATUS ret;
+
+	ret = validate_nodeid_array(&gpu_id_array,
+				NumberOfNodes, NodeArray);
+	if (ret != HSAKMT_STATUS_SUCCESS) {
+		free(gpu_id_array);
+		return ret;
+	}
+
+	ret = fmm_register_memory(MemoryAddress, MemorySizeInBytes,
+			gpu_id_array, NumberOfNodes);
+	if (ret != HSAKMT_STATUS_SUCCESS)
+		return ret;
+
+	return fmm_map_to_gpu(MemoryAddress,
+			MemorySizeInBytes, AlternateVAGPU);
+}
+
 HSAKMT_STATUS
 HSAKMTAPI
 hsaKmtUnmapMemoryToGPU(

src/topology.c

@@ -1317,6 +1317,30 @@ uint16_t get_device_id_by_gpu_id(HSAuint32 gpu_id)
 	return 0;
 }
 
+HSAKMT_STATUS validate_nodeid_array(uint32_t **gpu_id_array,
+		uint32_t NumberOfNodes, uint32_t *NodeArray)
+{
+	HSAKMT_STATUS ret;
+	unsigned int i;
+
+	if (NumberOfNodes == 0 || NodeArray == NULL || gpu_id_array == NULL)
+		return HSAKMT_STATUS_INVALID_PARAMETER;
+
+	/* Translate Node IDs to gpu_ids */
+	*gpu_id_array = malloc(NumberOfNodes * sizeof(uint32_t));
+	if (*gpu_id_array == NULL)
+		return HSAKMT_STATUS_NO_MEMORY;
+        for (i = 0; i < NumberOfNodes; i++) {
+		ret = validate_nodeid(NodeArray[i], gpu_id_array[i]);
+		if (ret != HSAKMT_STATUS_SUCCESS) {
+			free(gpu_id_array);
+			break;
+		}
+	}
+
+        return ret;
+}
+
 #if 0
 static int get_cpu_stepping(uint16_t* stepping)
 {