Adding support for new AQL Queue Memory allocation

Change-Id: If84fc4b961627dbdd0b77b1c509a3c9a4c709b9f
Signed-off-by: Ben Goz <ben.goz@amd.com>


[ROCm/ROCR-Runtime commit: e61500c46e]

projects/rocr-runtime/include/hsakmttypes.h

@@ -182,7 +182,8 @@ typedef union
         unsigned int DoorbellType        : 2;    // 0: This node has pre-1.0 doorbell characteristic
                                                  // 1: This node has 1.0 doorbell characteristic
                                                  // 2,3: reserved for future use
-        unsigned int Reserved            : 18;
+        unsigned int AQLQueueDoubleMap    : 1;	 // The unit needs a VA “double map”
+        unsigned int Reserved            : 17;
     } ui32;
 } HSA_CAPABILITY;
 
@@ -458,7 +459,10 @@ typedef struct _HsaMemFlags
                                            // 1: memory consistency needs to be enforced at
                                            // synchronization points at dispatch or other software
                                            // enforced synchronization boundaries.
-            unsigned int Reserved    : 18;
+            unsigned int AQLQueueMemory: 1; // default = 0; If 1: The caller indicates that the memory will be used as AQL queue memory.
+					    // The KFD will ensure that the memory returned is allocated in the optimal memory location
+					    // and optimal alignment requirements
+            unsigned int Reserved    : 17;
 
         } ui32;
         HSAuint32 Value;

projects/rocr-runtime/include/linux/kfd_ioctl.h

@@ -296,6 +296,8 @@ struct kfd_ioctl_set_process_dgpu_aperture_args {
 #define KFD_IOC_ALLOC_MEM_FLAGS_APU_DEVICE	(1 << 3)
 #define KFD_IOC_ALLOC_MEM_FLAGS_APU_SCRATCH	(1 << 4)
 
+#define KFD_IOC_ALLOC_MEM_FLAGS_DGPU_AQL_QUEUE_MEM		(1 << 5)
+
 struct kfd_ioctl_alloc_memory_of_gpu_new_args {
 	uint64_t va_addr;	/* to KFD */
 	uint64_t size;		/* to KFD */

projects/rocr-runtime/src/fmm.c

@@ -30,6 +30,7 @@
 #include <stdio.h>
 #include <inttypes.h>
 #include <sys/mman.h>
+#include <sys/time.h>
 
 #define NON_VALID_GPU_ID 0
 #define ARRAY_LEN(array) (sizeof(array) / sizeof(array[0]))
@@ -657,23 +658,30 @@ static void* fmm_allocate_host_gpu(uint32_t gpu_id,
 	manageble_aperture_t *aperture;
 	int32_t gpu_mem_id;
 	uint64_t mmap_offset;
+	uint32_t ioc_flags;
+	uint32_t size;
 
 	/* Retrieve gpu_mem id according to gpu_id */
 	gpu_mem_id = gpu_mem_find_by_gpu_id(gpu_id);
 	if (gpu_mem_id < 0)
 		return NULL;
 
+	size = MemorySizeInBytes;
+	ioc_flags = KFD_IOC_ALLOC_MEM_FLAGS_DGPU_HOST;
 	if (flags.ui32.CoarseGrain)
 		aperture = &gpu_mem[gpu_mem_id].dgpu_aperture;
 	else
 		aperture = &gpu_mem[gpu_mem_id].dgpu_alt_aperture; /* coherent */
+	if (flags.ui32.AQLQueueMemory) {
+		size = MemorySizeInBytes * 2;
+		ioc_flags = KFD_IOC_ALLOC_MEM_FLAGS_DGPU_AQL_QUEUE_MEM;
+	}
 
-	mem =  __fmm_allocate_device(gpu_id, MemorySizeInBytes,
+	mem =  __fmm_allocate_device(gpu_id, size,
 			aperture, 0, &mmap_offset,
-			KFD_IOC_ALLOC_MEM_FLAGS_DGPU_HOST);
+			ioc_flags);
 
 	/* FIXME: host memory allocated in this way should be mapped on all GPUs */
-
 	void *ret = mmap(mem, MemorySizeInBytes,
 			PROT_READ | PROT_WRITE | PROT_EXEC,
 		       MAP_SHARED | MAP_FIXED, kfd_fd , mmap_offset);