Align large buffers to BigK or huge-page boundary

This should allow us to take advantage of BigK fragments and huge pages
and improve TLB efficiency for VRAM allocations. Huge pages only work
with 4-level page tables (gfx900 and up). BigK fragments work on older
GPUs.

Change-Id: I02e1fbf74de554e16fdaf44e44d03b47df45c3b0
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>

src/fmm.c

@@ -539,11 +539,20 @@ static void *aperture_allocate_area_aligned(manageable_aperture_t *app,
 	vm_area_t *cur, *next;
 	void *start;
 
-	MemorySizeInBytes = vm_align_area_size(app, MemorySizeInBytes);
-
 	if (align < app->align)
 		align = app->align;
 
+	/* Huge-page and Big-K TLB optimizations require proper alignment */
+	if (MemorySizeInBytes >= GPU_HUGE_PAGE_SIZE) {
+		if (align < GPU_HUGE_PAGE_SIZE)
+			align = GPU_HUGE_PAGE_SIZE;
+	} else if (MemorySizeInBytes >= GPU_BIGK_PAGE_SIZE) {
+		if (align < GPU_BIGK_PAGE_SIZE)
+			align = GPU_BIGK_PAGE_SIZE;
+	}
+
+	MemorySizeInBytes = vm_align_area_size(app, MemorySizeInBytes);
+
 	/* Find a big enough "hole" in the address space */
 	cur = NULL;
 	next = app->vm_ranges;

src/libhsakmt.h

@@ -55,6 +55,12 @@ extern int PAGE_SHIFT;
 /* VI HW bug requires this virtual address alignment */
 #define TONGA_PAGE_SIZE 0x8000
 
+/* 64KB BigK fragment size for TLB efficiency */
+#define GPU_BIGK_PAGE_SIZE (1 << 16)
+
+/* 2MB huge page size for 4-level page tables on Vega10 and later GPUs */
+#define GPU_HUGE_PAGE_SIZE (2 << 20)
+
 #define CHECK_PAGE_MULTIPLE(x) \
 	do { if ((uint64_t)PORT_VPTR_TO_UINT64(x) % PAGE_SIZE) return HSAKMT_STATUS_INVALID_PARAMETER; } while(0)